Leveraging UAV hyperspectral imaging for crop physiology and biochemistry: A comprehensive review of feature extraction and selection methods

Abstract

Crop physiological and nutrient biochemical information plays a vital role in uncovering patterns of crop growth and development, as well as understanding their interactions with environmental factors. Unmanned aerial vehicles (UAV) -based hyperspectral imaging (HSI) technology offers an innovative tool for acquiring physiological and nutrient biochemical information through non-destructive and rapid collection of continuous spectral data from crops. However, challenges such as low signal-to-noise ratios (SNR), spectral variability for the same material, and high dimensionality in hyperspectral data make feature selection and extraction critical steps in data processing and analysis. Therefore, this review focuses on feature selection and extraction methods in the application of UAV-based hyperspectral technology for retrieving and monitoring crop physiological and biochemical information, providing theoretical support for its use in agriculture. Firstly, it provides a detailed discussion of feature selection methods, including filter-based, wrapper-based, and embedded approaches, along with various feature extraction techniques, analyzing their applicability and limitations in crop retrieving and monitoring. Secondly, the review highlights the use of vegetation indices (VIs) in feature extraction, covering advancements from basic indices to those optimized for specific applications. Finally, the article summarizes the main challenges of existing methods, particularly the issues of high-dimensional data processing and noise, and outlines potential future directions. This review highlights the significance of feature selection and extraction methods as critical tools for efficiently processing hyperspectral data. Through systematic analysis and synthesis, it provides theoretical support for agricultural researchers and practitioners while underscoring the importance of these techniques in driving innovation and advancements in hyperspectral technology.

1. Introduction

The inversion and estimation of crop physiological and biochemical information is a critical research area in agricultural science, directly linked to crop health, yield, and quality. Physiological parameters, such as biomass and leaf area index (LAI) [1], are key indicators of crop growth and health, directly influencing the photosynthetic activity and energy exchange [2,3]. Biochemical information primarily includes pigments (e.g., chlorophyll, carotenoids, and anthocyanins) and critical nutrient elements such as nitrogen, phosphorus, and potassium [4]. These biochemical parameters reflect crop photosynthetic efficiency and nutritional status. The traditional methods for evaluating crop traits are accurate but often destructive and time-consuming. For example, estimating AGB requires physical harvesting and weighing [5], while chemical analysis methods, such as the ethanol extraction of chlorophyll [6] and the Kjeldahl nitrogen method [7], are complex and destructive. Addressing these challenges, spectral imaging technology, particularly HSI, offers a revolutionary solution for crop trait monitoring. This technology enables non-destructive, rapid acquisition of spectral reflectance across continuous bands, allowing real-time access to physiological and biochemical information. Compared to traditional sensors, it provides more detailed spectral information [8], enabling researchers to precisely analyze the relationship between crop conditions and spectral data. With the advancement of UAV technology, agricultural remote sensing has entered a new era. UAVs equipped with hyperspectral camera systems overcome the limitations of ground-based and satellite remote sensing, enabling the rapid collection of detailed data over large areas. They significantly reduce the influence of weather conditions (e.g., solar angle, light intensity, wind, and clouds) on data collection [4], greatly enhancing the ability to monitor and invert crop physiological and biochemical traits. Overall, UAV-mounted hyperspectral technology provides robust support for efficient and dynamic monitoring of crop physiological and biochemical traits, advancing the development of precision agriculture.

Hyperspectral images captured by UAVs have nanometer-scale spectral resolution [9], enabling the detection of subtle spectral differences and providing highly precise spectral information. These images feature narrow and continuous spectral bands, covering the visible to near-infrared spectrum, which reflects an almost complete ground object spectral curve. In the inversion and monitoring of crop physiological and biochemical information, hyperspectral data provide abundant spectral features for identifying crop growth and nutritional status. However, hyperspectral images contain information from hundreds of continuous bands, resulting in extremely high feature dimensions for each pixel. This introduces significant computational and analytical challenges. First, noise in high-dimensional data and the high inter-band correlation often lead to information redundancy, increasing data processing complexity [10]. UAV hyperspectral images produce massive datasets with spatial, spectral, and temporal dimensions, further increasing the computational complexity of processing this data. While this high-dimensional feature space offers detailed crop information, it can also trigger the “Hughes phenomenon” [11], also known as the “curse of dimensionality.” The Hughes phenomenon refers to a decline in model performance due to increased noise and sparsity as the feature dimensions grow, leading to overfitting issues. When extracting physiological and biochemical indices, such as chlorophyll and nitrogen content, excessive spectral features may introduce redundant information and noise, disrupting the model's learning process and impairing the accuracy of crop status assessments. Moreover, the high inter-band correlation and collinearity lead to information redundancy, increasing computational demands and potentially reducing the model's generalization ability [12]. These issues pose significant challenges to the inversion and prediction studies of crop physiological and biochemical information. Therefore, in the agricultural application of UAV hyperspectral technology, effectively reducing data dimensionality and extracting sensitive features relevant to crop parameters has become a critical issue to address. Addressing this issue not only helps reduce information redundancy and enhance model performance but also provides more reliable technical means for precise monitoring and management of crop conditions.

To address the high dimensionality of hyperspectral data, dimensionality reduction has become essential. The primary goal is to represent high-dimensional data effectively with low-dimensional data, thereby reducing redundancy and improving processing efficiency. In hyperspectral data processing, dimensionality reduction methods can be broadly classified into feature selection and feature extraction [13]. Feature selection, also known as band selection, involves directly selecting a subset of spectral bands with high information content and low correlation from the original data. It retains the physical meaning of the spectral channels, facilitating interpretation and application while helping uncover underlying patterns and mechanisms. Feature selection methods can be categorized into filter-based, wrapper-based, and embedded approaches based on their relationship with model training [12]. Filter-based methods perform band selection before the model training, independent of the model. Wrapper-based methods select subsets of features and evaluate them using a model to identify the optimal bands. Embedded methods integrate feature selection into the model training process itself. Additionally, feature selection methods can be classified based on strategies such as search, ranking, clustering, and sparse theory. These methods are further divided into supervised and unsupervised approaches based on the use of training samples [14]. Common band selection methods include the successive projections algorithm (SPA) [15], competitive adaptive reweighted sampling (CARS) [16], recursive feature elimination (RFE) [17], VIP [18], and correlation analysis using Pearson's correlation coefficient [19]. Although band selection retains the physical meaning of the data, it has limitations. High-information bands are often highly correlated, making it challenging to meet the criteria of high information and low correlation simultaneously, which may limit the effectiveness of selected band combinations in practical applications. Additionally, in the presence of highly nonlinear data and large datasets, band selection algorithms are often complex, time-consuming, and inefficient. Feature extraction, on the other hand, uses mathematical transformations (e.g., PCA [20]) to map high-dimensional data to a new low-dimensional space. These methods integrate information from all bands, potentially improving data separability. However, if the transformation criteria are not chosen appropriately, the results may become difficult to interpret, and the physical meaning of the original bands may be lost. Moreover, vegetation indices are often used for feature extraction as they extract characteristics closely related to vegetation conditions from complex spectral information. Fig. 1 provides detailed information on methods for feature selection, feature extraction, and vegetation indices, presenting a comprehensive overview of methodological approaches, categorized into three main areas, each with detailed techniques and workflows. The feature selection section outlines widely used strategies, including filter, wrapper, and embedded methods, accompanied by representative techniques. The feature extraction section encompasses various approaches, such as clustering, spectral analysis, wavelet analysis, deep learning, and statistical methods. The vegetation index section summarizes basic indices, indices designed to eliminate interference factors, and those related to crop physiological and biochemical parameters, highlighting both systematic structure and practical applicability. Addressing these challenges, reasonable feature extraction and selection methods are critical in UAV hyperspectral technology applications. These methods not only effectively reduce data dimensionality, minimize information redundancy, and save computational resources but also improve model accuracy and stability, avoiding the curse of dimensionality [10]. In conclusion, adopting appropriate dimensionality reduction methods allows researchers to extract meaningful features for the inversion of crop physiological and biochemical information from extensive hyperspectral bands, enhancing monitoring accuracy and efficiency.

Fig. 1.

Overview of Feature Selection and Extraction Methods. The figure systematically illustrates three key methodological pathways in UAV hyperspectral remote sensing: Feature Selection, Feature Extraction, and Vegetation Index (VI) construction. Feature Selection is categorized into three mainstream strategies—Filter (A), Wrapper (B), and Embedded (C)—each illustrated above with schematic diagrams and accompanied below by representative algorithms. Feature Extraction includes clustering approaches (a), spectral analysis methods (b), wavelet transforms (c), deep learning techniques (d), and statistical methods (e). The Vegetation Index section summarizes basic indices (f), indices designed to reduce interference from soil, atmosphere, and other environmental factors (g), and indices closely related to crop physiological and biochemical parameters (h), such as those reflecting nitrogen content, pigment composition, structural properties, and biomass.

Given the above context, existing reviews primarily focus on either feature selection or feature extraction or are limited to discussions on hyperspectral image classification. There is a lack of comprehensive reviews on feature extraction and selection methods for UAV hyperspectral technology in the inversion and monitoring of crop physiological and biochemical information. Therefore, this study aims to fill this gap by systematically reviewing and analyzing feature extraction and selection methods for UAV hyperspectral technology in the inversion of crop physiological and biochemical information. We provide a detailed overview of existing band selection and feature extraction techniques, with a particular focus on the application of vegetation indices in feature extraction. By analyzing the effectiveness, applicability, and challenges of these methods in agricultural remote sensing, we explore how to maximize the extraction of sensitive features related to crop parameters while preserving the physical meaning of hyperspectral data. The structure of this paper is as follows: Chapter 2 outlines the literature search strategies, inclusion criteria, and statistical analysis of the included studies. Chapter 3 first clarifies the core concepts of dimensionality reduction, feature selection, and feature extraction, and subsequently elaborates on the complete preprocessing workflow for UAV hyperspectral data, ranging from fundamental corrections to signal optimization. Chapter 4 provides a detailed explanation of band feature selection methods and discusses filter-based, wrapper-based, and embedded approaches in relation to model training. Chapter 5 explores the main methods for band feature extraction. Chapter 6 focuses on feature extraction methods using vegetation indices, analyzing basic vegetation indices as well as those optimized for specific purposes. Chapter 7 summarizes the main challenges facing current feature selection and extraction methods and provides an outlook on future research directions. The final chapter concludes the study.

2. Materials and methods

We conducted a systematic review of published literature, focusing on the feature extraction and selection methods used for estimating crop physiological and biochemical parameters with UAV-based hyperspectral data. During the literature selection process, we adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines proposed by David Moher et al. [21], ensuring comprehensiveness and methodological rigor. The PRISMA framework comprises four key stages: identification, screening, eligibility assessment, and inclusion. This section provides a detailed description of the specific methods employed in the literature selection process for this review.

2.1. Literature search strategy

This study aims to analyze the feature extraction and selection methods used in UAV-based hyperspectral technology for retrieving and monitoring crop physiological and biochemical parameters. To achieve this, we performed a systematic search in the Web of Science [22] database, covering publications from 2014 to 2024. Web of Science is a comprehensive academic database that spans multiple disciplines, including natural sciences, engineering, and agriculture, and is widely used for the collection and analysis of scientific literature. Our search strategy utilized Boolean operators (“AND” and “OR “) in combination with carefully selected keywords to ensure both the precision and breadth of the search results. To align the search results with the research objectives, we refined the keyword combinations, focusing on terms such as “UAV “, “hyperspectral imaging “, “feature selection” or “feature extraction” along with the crop-related physiological and biochemical parameters, including pigments (e.g., “chlorophyll “), nutrients (e.g., “nitrogen “, “phosphorus “, “potassium “), “biomass “, “LAI” and “yield “. To enhance the relevance of the search, we refined the results within the “Agriculture” category under the Research Areas option in Web of Science, ensuring the retrieved literature closely aligns with the study's objectives. Table 1 lists the complete set of keywords and their classifications, which were combined across categories using Boolean operators to ensure the comprehensiveness and systematic nature of the search.

Table 1.

Keywords used during the literature search process for this review.

Category	Keywords (TS = )
Data Collection Method	(UAV OR “unmanned aerial vehicle”) AND (hyperspectral OR “hyperspectral imaging”)
Data Processing	(“feature selection” OR “feature extraction” OR “band selection” OR “band extraction”)
Applications	(chlorophyll OR carotenoids OR anthocyanins), (nitrogen OR phosphorus OR potassium), biomass, (“leaf area index” OR “LAI”), yield, (inversion OR monitoring OR prediction)
Research Area (Refine)	Agriculture

2.2. Process and criteria for article selection

This study aims to analyze the application of UAV-based hyperspectral technology in retrieving and monitoring crop physiological and biochemical parameters through a systematic literature review. Our selection criteria are strictly confined to agricultural crops, including wheat, maize, and soybean, with a focus on monitoring and retrieving key parameters such as nutrients (e.g., nitrogen, phosphorus, potassium), pigments (e.g., chlorophyll, carotenoids, anthocyanins), biomass, yield, and LAI. Studies on crop disease or pest classification are excluded from the scope of this review. The review covers studies published between 2014 and 2024, with no language restrictions, including publications in English and Chinese. To ensure the comprehensiveness and relevance of the review, we established detailed inclusion and exclusion criteria, as summarized in Table 2. This study includes research employing UAV-mounted hyperspectral sensors or handheld hyperspectral devices for retrieving and monitoring physiological and biochemical information in agricultural crops, with a particular focus on key parameters such as nutrients, pigments, biomass, yield, and LAI. Studies unrelated to agricultural applications, such as those focusing on forests or natural ecosystems, as well as research on disease and pest classification, are excluded from this review.

Table 2.

Selection criteria for the review process.

Category	Inclusion Criteria	Exclusion Criteria
Data Collection	Studies utilizing UAV-mounted hyperspectral sensors or UAVs combined with handheld hyperspectral devices for field data collection.	Studies relying solely on laboratory HSI systems or lacking relevance to field applications.
Study Subject	Research focused on agricultural or economic crops, including but not limited to wheat, maize, and soybean.	Research involving forests, natural ecosystems, or other non-agricultural applications.
Research Focus	Studies on crop physiological and biochemical information, such as pigments (e.g., chlorophyll), nutrients (e.g., nitrogen, phosphorus, potassium), biomass, LAI, and yield monitoring or retrieval.	Studies on crop disease or pest classification and monitoring.

In this study, we began the literature search by combining the keywords listed in Table 1, retrieving 256 articles from the Web of Science database. During the initial screening, we removed two duplicate records and excluded eight articles that fell outside the scope of the study, leaving 246 articles for further evaluation. In the subsequent screening, we excluded five review articles and obtained the full text of the remaining 241 articles for detailed analysis. However, seven articles were excluded due to inaccessibility. At the eligibility assessment stage, we rigorously evaluated 234 articles based on the predefined inclusion and exclusion criteria. Excluded articles primarily fell into the following categories: failure to meet UAV-based hyperspectral technology application standards, absence of relevant feature extraction or selection methods, or a research focus on forests, natural ecosystems, or non-agricultural pest and disease studies. Ultimately, 180 articles meeting the inclusion criteria were included in the review, 12 of which were published in Chinese. The detailed process of screening and evaluation is illustrated in Fig. 2 This literature selection process strictly adhered to the established research scope and criteria, ensuring the relevance and scientific rigor of the included articles to support the practical implications of the study.

Fig. 2.

PRISMA flow diagram of the literature selection process. The figure illustrates the PRISMA-compliant workflow for literature screening. A total of 256 records were retrieved from the Web of Science database. After duplicate removal, screening, and eligibility assessment, 180 studies were included in the final review. All included studies focus on crop monitoring using UAV-based hyperspectral technology. Excluded records consisted of review articles, studies unrelated to crops, and those focusing solely on pest or disease stress, ensuring consistency and relevance in the selected literature.

2.3. Quantitative analysis of the selected articles

An analysis of the 180 selected articles revealed a significant overall increase in the application of UAV-based hyperspectral imaging technology in agricultural crops over the past decade (Fig. 3 (a)). From 2014 to 2017, the number of related studies remained low, indicating that the field was still in its exploratory phase. Since 2018, the number of publications has steadily increased, although a slight decline in 2019 could be attributed to research limitations caused by the COVID-19 pandemic. However, the pandemic had a relatively minor impact as the application of this technology was still in its nascent stage. Starting in 2020, the number of publications grew rapidly, peaking in 2023 with 51 articles published. This remarkable growth is primarily attributed to the rapid advancements in UAV platforms and HSI technology, particularly in applications such as estimating crop physiological and biochemical parameters, detecting nutrient content, and predicting crop yields in precision agriculture. Hyperspectral technology captures spectral characteristics of crops that are beyond the reach of traditional imaging techniques, playing a crucial role in crop management and precision agriculture. As of October 2024, 21 articles have been published, and the annual number of publications is expected to decline slightly, likely reflecting the field's gradual maturation and stabilization rather than a decrease in interest. In summary, UAV-based hyperspectral technology has become an essential tool for crop monitoring and management, offering vast potential for future applications. This review highlights the indispensable role of hyperspectral feature extraction and selection in advancing this field.

Fig. 3.

The figure presents a statistical overview of research trends and key methodological approaches in UAV-based hyperspectral studies on crops. Subfigure (a) illustrates the annual publication trends from 2014 to 2024, highlighting a marked increase in research interest since 2020, with a peak observed in 2023. Subfigure (b) summarizes the frequency of individual and combined use of three core techniques: feature extraction, feature selection, and vegetation indices. Among them, vegetation indices are the most widely adopted. Subfigure (c) uses a ring chart to depict the distribution of three types of feature selection methods, with filter-based approaches accounting for the largest share (54.1 %). The outer ring provides a more detailed breakdown, showcasing representative algorithms such as correlation analysis (CC) and recursive feature elimination (RFE). Subfigure (d) displays the distribution of four categories of feature extraction techniques, indicating that statistical dimensionality reduction methods are currently the most dominant, representing 52.9 % of applications.

A systematic review of the existing literature reveals that feature selection, feature extraction, and vegetation indices are key methodologies commonly employed in remote sensing analysis. Statistical analysis shows that feature selection techniques were widely applied in 85 studies, while feature extraction methods were utilized in 34 studies. Additionally, vegetation indices, serving as a critical auxiliary analysis tool, were employed in 108 studies to improve the accuracy of classification and regression models. Many studies did not use these methods in isolation but combined them with other techniques, with the majority integrating vegetation indices. According to the statistics, 58 studies combined feature selection with vegetation indices, while 24 integrated feature extraction with vegetation indices to achieve the dual goals of dimensionality reduction and information enhancement. As shown in Fig. 3 (b), the frequency of application and the intersection of these methods in the literature underscore their significance and complementary advantages in remote sensing research.

Feature selection methods were categorized into three main types: filter-based, wrapper-based, and embedded methods. Statistical analysis indicated that filter-based methods were the most commonly used, appearing in 46 studies. Wrapper-based methods were the second most frequent, utilized in 27 studies, while embedded methods were the least common, mentioned in only 12 studies. Notably, most studies did not rely solely on one type of feature selection method but instead combined multiple approaches to optimize selection outcomes. Among filter-based methods, correlation analysis and the SPA were the most frequently used. For wrapper-based methods, RFE is the most commonly applied technique, closely followed by CARS. In embedded methods, commonly used techniques include feature selection from the model (SFM) and VIP. Fig. 3 (c) illustrates the specific usage proportions of each type of feature selection method. The “Other” category in the figure includes methods that appeared in only a few studies, reflecting their limited scope of application and lack of widespread adoption in literature.

Feature extraction methods can be broadly categorized into four main types. Statistical methods were the most frequently used, including techniques such as principal component analysis (PCA), partial least squares regression (PLSR), minimum noise fraction (MNF), independent component analysis (ICA), and non-negative matrix factorization (NMF). Spectral analysis methods ranked second, followed by clustering and wavelet-based methods. Additionally, some studies employed deep learning methods, which, despite their lower frequency of use, demonstrated potential in specific applications. Fig. 3 (d) illustrates the usage proportions of these four feature extraction categories, highlighting the dominance of statistical methods in this field.

We also conducted a statistical analysis of the applications of hyperspectral remote sensing in crop studies, as shown in Fig. 4.

Fig. 4.

Research Distribution of UAV Hyperspectral Technology Across Different Crop Domains. The bar chart presents the number of published studies involving various crop types in the context of UAV hyperspectral remote sensing. Wheat, rice, maize, and potato are the most frequently studied crops, highlighting their global agricultural significance and research priority.

Our findings indicate that research on UAV-based hyperspectral technology primarily focuses on major staple crops such as wheat, rice, maize, and potatoes. These crops have become key targets for hyperspectral remote sensing due to their extensive cultivation and critical role in global agricultural production. In these studies, the effective application of feature extraction and selection methods significantly improved the accuracy and efficiency of crop physiological and biochemical information retrieval and monitoring. For field crops such as wheat and maize, studies often utilize band selection methods to reduce redundant information in hyperspectral data and optimize monitoring of key parameters such as nitrogen content, LAI, and biomass. Additionally, using vegetation indices for feature extraction further enhances the accuracy of crop growth and health status estimation. For rice and potatoes, hyperspectral technology combined with feature extraction algorithms allows for more precise detection of rice chlorophyll content and potato tuber development, providing more accurate nutrient monitoring. Overall, these crop studies not only demonstrate the broad applicability of hyperspectral technology in agriculture but also highlight the critical role of feature extraction and selection methods in enhancing the accuracy of crop physiological and biochemical information retrieval and monitoring, particularly through the integration of band selection and vegetation indices for efficient data processing and application.

3. UAV hyperspectral data processing and analytical foundations

3.1. Conceptual clarification: dimensionality reduction, feature selection, and feature extraction

In hyperspectral remote sensing literature, the terms “dimensionality reduction,” “feature selection,” and “feature extraction” are often used interchangeably or ambiguously, which poses challenges to conceptual clarity. This section aims to clarify these core concepts and establish a clear terminological framework. The primary motivation for dimensionality reduction lies in addressing the inherent “curse of dimensionality” in hyperspectral data, as well as the resulting “Hughes phenomenon.” Dimensionality reduction seeks to process hyperspectral data efficiently while preserving as much relevant information as possible. This overarching goal is typically pursued through two main approaches: feature extraction and feature selection [13]. The fundamental distinction between the two lies in whether the physical meaning of the original spectral bands is retained. Feature selection identifies the most informative subset of original spectral bands, with its key advantage being the full preservation of physical interpretability—an essential aspect for understanding underlying mechanisms. Feature extraction, by contrast, applies mathematical transformations to map the original data into a new low-dimensional space. While this often sacrifices direct physical interpretability, it enhances the ability to capture complex underlying patterns and may lead to improved predictive performance. It is worth noting that although vegetation indices (VIs) are technically a form of feature extraction, they represent a special case designed with domain knowledge and thus retain interpretability. A detailed comparison of the two approaches is provided in Table 3.

Table 3.

Comparison between feature selection and feature extraction in UAV-Based hyperspectral applications.

Aspect	Feature Selection	Feature Extraction
Core Operation	To select	To transform/create
Output Features	A subset of the original features	New features derived from combinations of original features
Physical Interpretability	Preserved. Each selected feature retains its original physical meaning (e.g., reflectance at a specific wavelength).	Lost or obscured. Newly derived features cannot be directly interpreted as physical quantities.
Advantages	High interpretability; models are easier to understand and can be linked to physical processes; robust against irrelevant features.	Effectively captures latent data structures and enables efficient data compression; may yield higher performance in certain predictive tasks.
Limitations	Information loss due to feature elimination; may fail to capture complex interactions among features.	Lack of interpretability; performance may degrade when many irrelevant features are present.
Typical Methods	Filter, wrapper, and embedded methods (e.g., SPA, RFE)	Linear/nonlinear projections (e.g., PCA, autoencoders); domain knowledge-driven (e.g., vegetation indices)

3.2. Preprocessing of UAV-based hyperspectral data

Raw UAV-based hyperspectral data must undergo a series of preprocessing steps to eliminate noise and distortions introduced by the sensor, environmental conditions, and platform motion. This is a critical prerequisite for reliable downstream analysis [23]. The basic correction workflow generally involves radiometric and geometric corrections. Radiometric correction aims to convert raw digital number (DN) values into physically meaningful radiance or surface reflectance. In practice, this is commonly achieved using the empirical line method (ELM) based on field reference panels[[24], [25], [26]]. Some studies further incorporate fine-scale adjustments such as dark current correction [27,28]. Geometric and orthorectification processes are designed to correct spatial distortions caused by platform orientation and terrain effects, establishing a spatial reference for multi-temporal analysis and multi-source data integration. In practical applications, this step is often performed using commercial software such as Agisoft Metashape, in combination with high-precision IMU/GNSS data and ground control points (GCPs), to correct both geometric and optical distortions resulting from platform orientation and topographic variation[[29], [30], [31]]. Notably, the necessity of atmospheric correction for UAV data remains a subject of ongoing debate. On the one hand, the relatively short atmospheric path in low-altitude flights reduces the influence of atmospheric effects. On the other hand, unstable platform motion, complex viewing geometries, and micro-topographic variations increase the complexity of preprocessing. Given these trade-offs, a common practice is to either disregard atmospheric effects as negligible or assume that ELM implicitly corrects the major atmospheric influences, thereby omitting this step from the preprocessing pipeline [32,33]. Ultimately, the decision to apply rigorous atmospheric correction depends on the specific requirements for absolute accuracy and cross-temporal consistency in a given study.

Beyond basic corrections, the primary challenge lies in addressing the intrinsic quality of spectral signals and mitigating complex external interferences. To overcome low signal-to-noise ratios (SNR), spectral denoising and smoothing are the first essential steps. Among these, Savitzky–Golay (SG) filtering has become the most widely adopted benchmark method, as it effectively reduces noise while preserving spectral features [34]. It has been applied in studies of rice, cotton, potato, and other crops [24,25,27,35,36]. In addition to SG filtering, other algorithms such as Gaussian filtering, mean filtering, and locally weighted scatterplot smoothing (Lowess) regression have also been employed [31,[37], [38], [39]]. A more direct cleaning strategy involves removing edge bands with extremely low SNR [40,41]. Model-driven approaches such as minimum noise fraction (MNF) represent more advanced denoising trends [2,42]. After smoothing, derivative spectroscopy (including first-order and fractional derivatives) is widely used to correct spectral variations caused by sample properties and illumination changes. This method reduces baseline drift and amplifies subtle hidden spectral signals [33,43,44]. Continuum removal, by normalizing absorption features, further enhances the comparability of spectra [45]. Additionally, techniques such as standard normal variate (SNV) transformation and multiplicative scatter correction (MSC) are employed to reduce physical scattering effects, effectively improving spectral stability in crops such as cotton and maize [35,46]. Finally, before feeding the data into machine learning models, numerical scaling such as Z-score standardization or Min–Max normalization is typically performed to meet the training requirements of specific algorithms [47,48].

In agricultural scenarios with complex background heterogeneity, background removal is a critical step in data preprocessing. A straightforward strategy is to apply thresholding based on vegetation properties, such as high reflectance in the near-infrared (NIR) band or vegetation indices (e.g., the Excess Green Index, ExG) [38,[49], [50], [51]]. When simple thresholding fails to address complex field conditions, machine learning classifiers such as Support Vector Machines (SVMs), or advanced deep learning models such as Mask R-CNN, are often employed for more precise segmentation [28,42]. However, a major challenge in hyperspectral imaging is that spectral information varies spatially, resulting in mixed pixels in which a single pixel represents multiple ground features [52]. The fundamental technique to address the mixed pixel problem is spectral unmixing. This method aims to mechanistically decompose pixel signals into pure components (endmembers) such as vegetation, soil, and shadow, along with their corresponding abundances. For instance, in precision fertilization studies on rice, Yu et al. [42] applied the Pixel Purity Index (PPI) algorithm to extract endmembers. Unlike approaches that directly adjust spectral curves, spectral unmixing transforms high-dimensional spectral data into a small set of abundance values with clear physical meaning (e.g., crop canopy coverage). These abundance values serve as condensed macro-features, offering a sharp contrast to strategies that simply select sensitive spectral bands. This represents a distinct, model-driven pathway for feature generation. Therefore, spectral unmixing is not only an advanced data purification technique but also a unique paradigm for feature extraction. Specific algorithms for implementing this approach, such as Non-negative Matrix Factorization (NMF), will be discussed in detail in the feature extraction section. From the perspective of feature engineering, the significance of spectral unmixing is profound, as it represents a distinct, model-driven pathway for generating features. In summary, preprocessing steps ranging from signal optimization to background removal collectively determine the upper bound of performance for subsequent feature selection and extraction algorithms.

4. Band feature selection methods

In the application of UAV-based hyperspectral imaging technology, band feature selection plays a pivotal role in determining the accuracy of crop retrieval and monitoring. The primary goal of feature selection is to identify the most relevant spectral bands from high-dimensional data, thereby reducing data dimensionality, eliminating redundant information, and enhancing model efficiency and accuracy. As shown in Fig. 5, the process of band feature selection does not involve transforming the original bands.

Fig. 5.

Workflow of Hyperspectral Band Feature Selection. The figure illustrates the typical process of selecting informative bands from original hyperspectral data. Various feature selection algorithms—such as Correlation Coefficient (CC), Successive Projections Algorithm (SPA), Variable Importance in Projection (VIP), and Recursive Feature Elimination (RFE)—are applied to eliminate redundant or less informative bands. The selected subset (e.g., {b₁, b₄, b_i−2, b_i}) retains the most relevant spectral features for modeling, improving both model stability and generalization without altering the physical meaning of the original bands.

Instead, it evaluates the correlation between each band and the target variable using various methods, selecting the most representative and informative bands for subsequent research. Feature selection methods are typically categorized into three types: filter methods, wrapper methods, and embedded methods [12]. Filter methods assess the relevance of features to the target variable using statistical metrics, wrapper methods evaluate subsets of features based on learning algorithms, and embedded methods perform feature selection concurrently with model construction. Each of these methods has distinct advantages, which will be discussed in detail in the following subsections.

4.1. Filter methods

Filter-based methods are a widely used strategy in feature selection, relying on statistical metrics or scoring criteria to independently evaluate each feature without depending on specific machine learning algorithms. These methods are typically employed as a preprocessing step before model training, aiming to assess the relationship between features and target variables through statistical analysis [53]. Due to its computational efficiency and intuitive principles, this method is well suited for preliminary feature screening in large-scale datasets. However, because it evaluates features independently, it may overlook complex relationships among them. Nevertheless, filter-based methods remain important in feature selection due to their simplicity and efficiency, providing effective support for subsequent model development.

Among filter-based feature selection methods, Correlation Analysis (CC) and the Successive Projections Algorithm (SPA) are two of the most classic and complementary approaches. Their core principles are “maximizing correlation” and “minimizing collinearity,” respectively. Together, they reflect the essential trade-off between efficiency and redundancy inherent in this class of methods. Both methods are widely applied in UAV-based hyperspectral (UAV-HSI) monitoring of crop physiological and biochemical parameters, and they are often combined to balance efficiency and accuracy. CC, as a fundamental and critical feature selection method, operates by maximizing the association between features and the target variable. It uses indicators such as the Pearson correlation coefficient to rapidly assess the linear or monotonic relationships between each spectral band and crop parameters, making it one of the most direct and computationally efficient tools for identifying sensitive bands. In some cases, simple CC-based screening is sufficient to build high-precision models. For example, Zhao et al. [37] combined spectral data with vegetation height and accumulated temperature, then used CC to select features significantly correlated with summer maize LAI and SPAD (Pearson R $>$ 0.5). These features were directly applied in modeling, yielding highly accurate estimates of LAI (R² = 0.93). However, CC has inherent limitations due to its independent evaluation mechanism, leading to two main issues: first, its inability to capture complex nonlinear relationships; and second, more critically, its complete neglect of the multicollinearity commonly present in hyperspectral data. As a result, while the selected features may each be strongly correlated with the target, they are also often highly correlated with each other, which often leads to substantial redundancy. To address this limitation, researchers have developed modified correlation analysis approaches that consider both “feature–target” relationships and “feature–feature” relationships. Common practices include using spectral autocorrelation matrices [32] or setting explicit correlation thresholds [33,54] (e.g., r $>$ 0.95) to actively eliminate redundant bands. This straightforward modification has proven to be highly effective. For instance, Liu et al. [32] used a spectral autocorrelation matrix to select representative bands. With only 4–6 selected bands, they constructed a high-accuracy model that achieved an R² of 0.969 on independent-year validation data, demonstrating that a small number of carefully selected, non-redundant bands can effectively capture key physiological information. In another study, Liu et al. [33] applied a correlation-thresholding approach to estimate the nitrogen nutrition index (NNI) in winter wheat. Their multivariate linear model achieved good accuracy in independent validation (R² = 0.817 at flowering). Similarly, Zhu et al. [54] used a threshold-based method to construct a new vegetation index from selected bands, demonstrating that model accuracy improved by up to 45 % compared with traditional indices. This idea was also applied by Li et al. [55] in estimating nitrogen content in apple canopies. They proposed a Modified Correlation Coefficient Method (MCCM), which similarly employed an autocorrelation threshold to ensure low redundancy and high representativeness in the final feature set. Collectively, these findings demonstrate that the effective removal of redundant information not only preserves key predictive power but also enhances model stability and generalization by reducing complexity and noise.

Complementary to CC, the Successive Projections Algorithm (SPA) is explicitly designed to minimize collinearity [15]. Instead of evaluating each band's independent relationship with the target, SPA employs a forward-selection geometric projection mechanism that ensures newly selected bands have the lowest linear correlation with those already chosen. This process produces smaller yet more representative feature subsets. Owing to its strong capability for redundancy reduction, SPA has been widely applied in UAV-based hyperspectral feature selection for retrieving physiological parameters in crops such as rice, potato, tobacco, and licorice [24,27,38,42,45,50,56,57]. A notable trend is that SPA has evolved from being merely paired with traditional regression models to serving as an indispensable preprocessing step for advanced machine learning models. Studies have shown that integrating SPA with Extreme Learning Machines (ELMs), optimized by metaheuristic algorithms, significantly enhances both predictive accuracy and stability compared with baseline models [24,38,42]. The work of Zhang et al. [50] further highlights SPA's value. When features selected by SPA were input into a Stacking ensemble model, the estimation of tobacco nitrogen content improved in accuracy (validation R² increased from 0.68 to 0.745) and demonstrated stronger robustness in cross-regional validation compared with single models such as PLSR.

CC ensures the individual validity of features, while SPA secures the overall efficiency of the feature set. This complementarity provides the theoretical foundation for more sophisticated combined strategies. To overcome the limitations of single filter methods such as CC and SPA, an efficient paradigm known as the “coarse-to-fine” strategy has been widely adopted in practice. This framework integrates the strengths of different methods to balance feature selection performance and computational cost, and has become a well-established standard workflow. The core mechanism is to first take advantage of CC's high computational speed for preliminary screening, rapidly eliminating clearly irrelevant or highly redundant variables from hundreds of raw bands or large candidate features (e.g., using thresholds such as |r| > 0.95 or 0.98 [29,58]). This step produces a substantially reduced and more manageable feature space for subsequent computationally intensive “fine selection” algorithms. Next, more computationally demanding but precise methods—such as wrapper approaches (e.g., RFE, MSR) or embedded approaches (e.g., XGBoost)—iteratively optimize the pre-screened feature set, capturing feature interactions overlooked by CC. The effectiveness of this hybrid strategy has been widely validated across diverse crops and parameter estimation tasks, consistently revealing a clear hierarchy of performance. Comparative studies clearly demonstrate that the subsequent “fine selection” step is crucial: wrapper methods such as Boruta and RFE, as well as more sophisticated regression models (e.g., GPR-SBBR), consistently outperform CC-based baseline models. The best combined models can achieve predictive accuracies as high as R² = 0.96 [51,59]. Specific case studies further confirm this point. For instance, researchers successfully combined CC with RFE to estimate maize LAI (R² = 0.89) [58]; integrated CC with MSR to monitor winter wheat SPAD (R² = 0.85 at flowering) [60]; and coupled CC with XGBoost and deep neural networks (DNNs) to retrieve rice LNC (R² = 0.89) [31]. In each case, performance significantly exceeded that of single methods.

This paradigm also demonstrates remarkable flexibility, making it well suited for handling complex datasets composed of multi-source and heterogeneous features. Among these are vegetation indices (VIs), an important class of features whose selection, construction, and application will be discussed in detail in Chapter 6. For example, when processing a large feature pool that includes raw spectra, spectral transformations, positional parameters, and VIs, Liu et al. [25] first applied CC for coarse screening, followed by systematic comparisons of advanced algorithms such as MC-UVE and Random Frog, ultimately identifying the optimal combination. Similarly, this strategy can even integrate feature selection (CC) with feature extraction (PCA) to achieve deeper dimensionality reduction and decorrelation. In estimating Chinese milk vetch biomass, Hu et al. [47] achieved strong performance using this approach (R² up to 0.813). The workflow can also be further refined, such as in the three-step procedure of “CC coarse screening → SPA collinearity removal → PLSR weight-based selection,” which aims to achieve optimal filtering performance [45]. Thus, the “coarse-to-fine” combination strategy represents an effective paradigm for balancing computational efficiency and model accuracy. However, there is no universally optimal combination; the choice of strategy must be tightly coupled with the specific agricultural application context. For instance, Li et al. [57] found in their study of potato chlorophyll that although SPA was applied consistently across all growth stages, the best-performing predictive model varied with developmental stage. This conclusion was reinforced by Xu et al. [27] in a comprehensive comparison study on licorice. By directly comparing SPA, CARS, GA, and RFE, they showed that during flowering the SPA-PLSR combination performed best (R² = 0.77), whereas during regreening the GA-XGBoost combination achieved the highest accuracy (R² = 0.95). These findings strongly indicate that the optimal feature selection strategy is not static but must be dynamically adjusted according to crop phenological stages to account for the changing relationships between spectral signals and physiological parameters. Collectively, these studies suggest that sacrificing a small amount of initial computational efficiency in exchange for subsequent exploration of superior feature subsets using advanced algorithms is a judicious choice in current research.

Beyond basic correlation and collinearity analyses, a variety of filter-based strategies enrich the feature selection toolbox by introducing perspectives that emphasize feature interactions, robustness to noise, and even physical mechanisms. Among these, the ReliefF algorithm serves as an important bridge between simple univariate evaluations and more complex wrapper methods. Unlike CC or SPA, which prioritize computational speed, ReliefF evaluates the ability of each feature to distinguish between near-neighbor samples of the same and different classes. This approach introduces sensitivity to complex feature interactions without substantially increasing computational cost [61]. Consequently, ReliefF is often integrated into ensemble selection frameworks to provide a unique evaluation perspective. For example, in peanut water-stress classification, it achieved an accuracy of 96.46 % [48]. Its regression variant, RReliefF, demonstrates particular advantages in handling continuous variables and heterogeneous features. A typical application is provided by Li et al. [62], who applied RReliefF not to raw spectra but to engineered vegetation indices (VIs). Their potato biomass regression model achieved R² = 0.92, highlighting the algorithm's strength in evaluating and integrating multi-source information. Another class of methods, such as Uninformative Variable Elimination (UVE) and its Monte Carlo variant (MC-UVE), introduces random noise variables to more reliably remove bands with unstable coefficients. These methods achieved validation R² values of 0.887 for rice SPAD prediction and 0.89 for potato biomass prediction [25,63]. In addition, classical statistical approaches such as the chi-square test and Fisher Score offer a foundational evaluation perspective. However, since they assess features only in a univariate manner and ignore feature interactions, their standalone performance is limited. A comparative study clearly confirmed this limitation: an SVM model built on features selected using the chi-square test achieved only 83.08 % accuracy, far lower than the 95.38 % accuracy of RFE-SVM [36]. Therefore, their role in contemporary research is shifting from standalone filters to supplementary components within ensemble selection frameworks, where they contribute a basic statistical perspective and enhance the diversity of evaluation dimensions [48].

In sharp contrast to the data-driven methods discussed above, the PROSAIL model combined with lookup tables (LUTs) represents a physics-driven approach to feature selection. Its main advantage lies in selecting features based on the physical causal relationships derived from simulating photon radiative transfer within the canopy, rather than on statistical correlations. As a result, the selected features offer strong interpretability and broad theoretical generality [64]. However, its performance is highly dependent on how well its physical assumptions align with real-world conditions. A series of critical comparative studies have provided valuable quantitative evidence. Under real UAV observation conditions, data-driven models (e.g., Random Forest) often outperform PROSAIL-LUT due to their strong nonlinear fitting capabilities, achieving higher estimation accuracy (e.g., LAI R² = 0.83 vs. 0.77 for PROSAIL-LUT) [65]. The limitations are particularly evident in row-planted crops with discontinuous canopy structures. For example, PROSAIL's estimation errors for maize LAI reached an RRMSE of 31.1 %, much higher than for potato (5.2 %) and sunflower (10.3 %). Studies attribute this to the model's inability to account for inter-row shading [23]. Its performance is also highly sensitive to crop growth stages and soil background conditions. A study by Wang et al. [66] on rice provides a clear example: during the tillering stage, when canopies are sparse, PROSAIL-LUT produced the largest errors, with LAI being severely overestimated. The study also quantified the decisive role of background settings: only when the model used a flooded-soil background (reflecting real paddy field conditions) instead of a generic bare-soil background could it achieve accurate retrievals across all growth stages, with canopy chlorophyll content (CCC) and LAI estimation accuracies reaching R² = 0.79 and 0.70, respectively. Therefore, choosing PROSAIL-LUT essentially entails a fundamental trade-off between interpretability and theoretical generality, on the one hand, and the flexibility and maximal predictive accuracy of data-driven models, on the other.

In addition, other strategies have further extended the boundaries of filter-based methods. Some approaches introduce unique evaluation perspectives from information theory or geometric relationships. For example, Shannon entropy quantifies the information value of each band to select features, achieving validation accuracies of R² = 0.81 for grassland biomass and R² = 0.70 for nitrogen content estimation [67]. In geometry-based methods, Grey Relational Analysis (GRA) identifies the most relevant bands for winter wheat growth traits by evaluating sequence-shape similarity, achieving correlation coefficients up to r = 0.875 [68]. Similarly, multidimensional Euclidean distance directly measures the proximity between features and targets in multidimensional space, yielding an R² of 0.758 for rice grain protein content prediction [69]. Other, more complex strategies focus on system-level analyses. For instance, two-dimensional correlation spectroscopy (2D-COS) reveals deeper associations by analyzing spectral dynamics under external perturbations such as growth stage, significantly improving winter wheat chlorophyll estimation accuracy (R² = 0.85) [70]. Global Sensitivity Analysis (GSA), such as the Sobol method, quantifies the contribution of each input band to output variance in complex models. The identified key features supported the construction of high-accuracy models for LAI (R² = 0.90) and LCC (R² = 0.86) [71]. Grid search and similar approaches elevate feature selection to the level of feature design and optimization. For example, an optimized NDVI constructed through exhaustive search significantly outperformed all standard vegetation indices [72]. Together with classical methods such as the chi-square test, these approaches constitute a diverse toolkit that provides flexible options for handling complex UAV hyperspectral datasets.

In summary, filter-based methods form the foundation of hyperspectral feature selection, encompassing critical methodological trade-offs. Research practice has clearly shown a shift from searching for a single optimal method toward building multi-stage, high-efficiency combination strategies, with the “coarse-to-fine” paradigm as the mainstream approach. This transformation reflects two fundamental trade-offs: (i) within data-driven approaches, balancing the extreme speed of correlation analysis (CC) with the redundancy-reduction depth of SPA or ReliefF; and (ii) at a broader methodological level, weighing the predictive accuracy of data-driven models against the theoretical interpretability of physics-based models such as PROSAIL. This transformation reflects two fundamental trade-offs: (i) within data-driven approaches, balancing the extreme speed of correlation analysis (CC) with the redundancy-reduction depth of SPA or ReliefF; and (ii) at a broader methodological level, weighing the predictive accuracy of data-driven models against the theoretical interpretability of physics-based models such as PROSAIL. Therefore, the most important conclusion of this section is that no universally optimal filter strategy exists. The choice of method must be closely coupled with the specific agricultural context— including crop type, developmental stage, and monitoring objective—if high-accuracy and robust monitoring is to be achieved.

4.2. Wrapper methods

Unlike filter-based methods, wrapper methods are strategies that tightly integrates feature selection with model training to identify the optimal subset of features. These methods evaluate the effectiveness of each feature subset based on the model's training accuracy or other performance metrics, iteratively selecting features to determine the subset that maximizes model performance [53,61]. Unlike filter-based methods, wrapper methods are directly guided by model performance, leveraging feedback from the model to iteratively optimize feature selection. This approach fully accounts for feature interdependencies and complex relationships. This mechanism often yields superior feature sets and higher predictive accuracy. However, its high computational cost and susceptibility to overfitting pose major challenges when dealing with high-dimensional data [12]. Nevertheless, owing to its strong adaptability to specific tasks, the wrapper approach remains widely applied in hyperspectral feature selection for unmanned aerial vehicles.

The central challenge of wrapper methods lies in searching the feature space. The most intuitive and classical approach is deterministic greedy search, which makes a locally optimal choice at each step. Recursive Feature Elimination (RFE) is the most powerful and widely used representative of this strategy. Its core principle is to train a chosen learning model (e.g., support vector machines, random forests, etc.) iteratively. After each round, the importance of each feature is evaluated, and the least informative features are systematically removed until the model achieves optimal performance [41,51]. The optimization objective of RFE is the final predictive accuracy of the model. This sets it apart from other approaches: rather than pursuing mathematical optimality (i.e., minimum collinearity) as in the Successive Projection Algorithm (SPA), RFE seeks predictive optimality. Moreover, as a general framework that can be combined with any machine learning model, RFE offers exceptional flexibility. The strength of RFE lies in both its wide applicability and its strong task specificity. It can be applied to regression tasks (e.g., predicting alfalfa yield with R² = 0.874 [41]) as well as classification tasks (e.g., detecting water stress in pearl millet with an accuracy of 95.38 % [36]). More importantly, RFE yields highly task-specific feature selections, adaptively identifying the most relevant features depending on the crop type, growth stage, and target parameter. Zhu et al. [29] demonstrated this adaptive capability: across crops, RFE favors features at different scales (canopy features for wheat, leaf features for maize); within the same crop, it captures phenological differences—wheat achieved the best performance at flowering (R² = 0.79), whereas maize reached peak accuracy at the grain-filling stage (R² = 0.87). When the prediction target shifts from physiological parameters (LCC) to physical parameters (AGB), RFE can flexibly pivot to LiDAR or RGB texture features, highlighting its robustness under complex scenarios. To mitigate the high computational cost, researchers often adopt a two-step framework of “coarse screening + refined selection” [27,29], or they refine the algorithm itself. For example, Recursive Feature Elimination with Cross-Validation (RFECV) can effectively identify key bands even under suboptimal conditions such as crop lodging, further confirming the robustness of this method [73].

As a computationally lighter and more statistically interpretable variant of the greedy search paradigm, Multiple Stepwise Regression (MSR) has also been widely applied. Unlike RFE, which requires repeated training of complex models, MSR relies on linear models and statistical criteria (e.g., p-values from F-tests) to automatically add or remove variables. For instance, Wang et al. [74] used MSR to select the optimal combination of vegetation indices for estimating SPAD values in winter wheat, significantly improving predictive accuracy at the flowering stage (R² = 0.85). Similarly [75], employed MSR to identify an optimal set of bands, including 399 nm, 520 nm, and 668 nm. These served as high-quality, low-dimensional inputs for a more advanced Extreme Learning Machine (ELM) model, which ultimately achieved very high validation accuracy (R² = 0.968). In addition, MSR has proven effective in integrating heterogeneous feature sources [76]. demonstrated that MSR can successfully combine spectral and texture features to improve the estimation of above-ground biomass (AGB) in rice, achieving particularly high accuracy during the tillering stage when the canopy is sparse (R² up to 0.821). However, whether in the more complex RFE or the simpler MSR, both methods follow a greedy “step-by-step decision” logic with deterministic search paths, making them prone to local optima. When addressing complex problems involving highly nonlinear feature interactions, this limitation has motivated the development of more powerful global search strategies.

To overcome the limitation of deterministic searches that often fall into local optima, researchers have introduced metaheuristic algorithms, which mimic natural evolution or swarm behavior to achieve global optimization through randomization and parallel mechanisms. Common representatives include Genetic Algorithms (GA), Shuffled Frog-Leaping Algorithm (SFLA), and Random Frog. Although their underlying mechanisms differ—GA relies on evolutionary “survival of the fittest,” SFLA simulates frog colonies performing global and local searches, and Random Frog uses Reversible Jump Markov Chain Monte Carlo (RJMCMC) for stochastic sampling—their core idea is consistent: maintaining a population of candidate solutions and using selection, crossover, mutation, or iterative evaluation to explore the feature space and escape local traps. The value of these algorithms is most evident in handling complex, nonlinear relationships where context matters. For example, in estimating chlorophyll content in licorice, the combination of GA and XGBoost achieved the best performance only during the early growth stage (returning green), when relationships were highly complex, reaching an accuracy of R² = 0.95. In later stages, simpler methods such as SPA or CARS proved more effective [27]. At the same time, advanced heuristic methods such as SFLA have demonstrated clear superiority in direct performance comparisons. Ma et al. [35] showed that SFLA could identify band combinations with stronger synergistic effects than those obtained by SPA, improving the estimation accuracy of cotton LAI to R² = 0.907. Liu et al. [25] provided further evidence, showing that Random Frog significantly outperformed MWPLS, another wrapper method based on local window search. The fundamental reason lies in the inherent limitation of MWPLS's “local optimization,” which restricts its ability to identify globally optimal combinations. Taken together, metaheuristic algorithms show great potential in capturing complex feature interactions and achieving global optimization. They represent an important pathway toward maximizing predictive accuracy when computational resources allow.

Beyond the general search frameworks described above, several highly specialized feature selection tools have been developed, tightly coupled with specific regression models or theoretical foundations. Competitive Adaptive Reweighted Sampling (CARS) is a typical example. It evaluates band importance through repeated random sampling combined with Partial Least Squares Regression (PLSR). Although it borrows the “survival of the fittest” principle similar to GA, CARS is deeply dependent on the PLSR model, making it fundamentally different from the general frameworks discussed earlier. Unlike GA, which simulates biological evolution to evolve “feature subsets” as a general global optimization strategy, or RFE, which uses deterministic backward elimination and can wrap around any model, CARS is intrinsically bound to PLSR. At its core, CARS uses PLSR regression coefficients as the fitness criterion, making it particularly suited for handling highly collinear spectral data. Researchers often combine CARS with the Successive Projection Algorithm (SPA), forming the efficient “CARS + SPA” composite strategy [77]. SPA first eliminates collinearity, followed by CARS for performance-driven fine selection. The effectiveness of this combination stems from the superior performance of CARS over SPA. For example, Shu et al. [49] systematically compared the two and found that CARS-PLS outperformed SPA in estimating maize SPAD values (R² = 0.86) and total leaf area (R² = 0.62). Zhang et al. [78] applied this combination to identify nine optimal bands, improving the accuracy of winter wheat LAI estimation to R² = 0.89. Similarly, Sudu et al. [40] used the same strategy with a Deep Neural Network (DNN), raising the accuracy of summer maize SPAD estimation to R² = 0.82. However, the strength of CARS lies in its strong task specificity. Wang et al. [74] showed that it performed best in estimating wheat leaf nitrogen content (LNC, R² = 0.87), but was outperformed by Random Forest (RF) in predicting canopy above-ground biomass (AGB), which is influenced by multiple factors. This indicates that CARS is better suited for physiological and biochemical parameters, while it faces limitations in tasks involving complex biophysical parameters.

Another specialized tool is Gaussian Process Regression (GPR). As a non-parametric Bayesian method, it can directly assess feature importance through the Automatic Relevance Determination (ARD) kernel while simultaneously providing predictive uncertainty. For example, Verrelst et al. [79] developed the GPR-BAT tool, which combines ARD with Sequential Backward Band Removal (SBBR). Using only 4–9 key bands selected by GPR, the method achieved performance comparable to or exceeding full-spectrum models in estimating parameters such as chlorophyll and LAI (e.g., gLAI estimation with RCV² = 0.94), while significantly reducing computational cost. Similarly, Song et al. [59] applied a GPR + SBBR strategy to identify optimal combinations of vegetation indices (VIs) for estimating nitrogen status in winter wheat. This approach outperformed traditional MLR and parametric regression methods, achieving an R² as high as 0.96 for plant nitrogen concentration (PNC). However, a key limitation of GPR is its inherent computational complexity. With a cost of O(N³) relative to sample size N, directly handling high-dimensional data imposes a substantial computational burden. Wen et al. [80] offered a successful example of addressing this challenge. They first applied Principal Component Analysis (PCA) to extract key components, which were then used as inputs to the GPR model, yielding strong predictive accuracy (R² = 0.8525). Overall, CARS and GPR, as two powerful specialized tools, provide unique perspectives for feature selection. Yet, their use requires a careful balance between task-specific advantages, distinctive functionalities, and computational cost.

The inherent tension between the high computational cost and strong optimization capacity of wrapper methods naturally confines their practical use to the “coarse-to-fine” framework introduced earlier for filter-based methods. Within this framework, wrapper methods do not serve as stand-alone solutions; instead, they play the crucial role of refinement. Because of their computational complexity, applying RFE, GA, or similar methods directly to high-dimensional raw data is often infeasible. In practice, researchers typically begin with correlation analysis, SPA, or other efficient filter-based methods to rapidly reduce the feature space from hundreds to a few dozen, and then apply computationally intensive yet more precise wrapper methods (e.g., RFE, GA, CARS) for a second round of optimization. Hybrid strategies such as “CARS + SPA,” as well as preselection followed by RFE, have confirmed the feasibility of this approach. Essentially, this strategy trades a small amount of initial efficiency for deeper exploration of high-quality feature subsets, thereby achieving a balance between efficiency and performance.

In summary, the greatest value of wrapper methods lies in their feedback-driven approach, which can capture complex feature interactions and significantly enhance predictive accuracy. However, these advantages come at the cost of high computational demands and a risk of overfitting. A comparison of strategies shows that RFE represents efficient deterministic search, GA and other metaheuristic algorithms emphasize global optimization, while the performance of specialized tools such as CARS is highly dependent on the task context. Therefore, the optimal role of wrapper methods is not to be used independently, but rather as a critical component in the “filter-based screening followed by wrapper-based refinement” framework. By first narrowing the candidate feature space and then applying deep optimization, researchers can not only ensure computational feasibility but also maximize the potential of these methods.

4.3. Embedded methods

Embedded band selection methods perform feature selection automatically during the model-building process [81], effectively avoiding the need to repeatedly train learners, resulting in band selection outcomes that are better aligned with the model's adaptability. Embedded methods integrate learning and feature selection. Compared to wrapper methods, embedded methods do not require repeated evaluation and training for every feature subset, making them more computationally efficient. Compared to filter methods, embedded methods are better aligned with specific models, thereby improving overall model performance. However, the successful application of embedded methods also relies on careful objective function design and parameter optimization. In UAV hyperspectral analysis, based on their underlying model assumptions, commonly used embedded methods can be clearly divided into two major groups. These are methods based on linear frameworks and those based on nonlinear ensemble learning [82].

Methods based on linear frameworks are distinguished by their efficiency and interpretability. Among them, Variable Importance in Projection (VIP) is closely coupled with the Partial Least Squares Regression (PLSR) model, and its core value lies not only in feature selection but also in model interpretation. It quantifies the contribution of each spectral band to model explanation and prediction by computing a composite VIP score (commonly adopting a threshold of >1), thereby revealing the basis of model decisions. Compared with filter-based methods such as SPA, it incorporates more model information, which creates a unique trade-off between efficiency and flexibility. Unlike wrapper methods (e.g., RFE), it does not require costly iterative modeling, making it more computationally efficient. However, its limitation lies in its deep dependence on the PLSR model, sacrificing the high flexibility of RFE, which can wrap around any nonlinear model. Therefore, the greatest limitation of VIP is that its validity entirely depends on the explanatory power of PLSR; when the data relationships are highly nonlinear, its guidance becomes less meaningful. Despite this limitation, VIP remains a powerful tool for diagnosing and understanding PLSR models in appropriate contexts. Zhou et al. [83] used VIP to identify the most informative intervals from hundreds of bands, achieving high predictive accuracy (R² = 0.78) in estimating potato nitrogen concentration. Lin et al. [46] further demonstrated that VIP not only extracts key bands but also provides a stable basis for modeling different nutrients, with PLSR models achieving independent validation accuracies of R² = 0.70 for magnesium (Mg) and R² = 0.72 for potassium (K). These studies indicate that VIP is not only an efficient feature selection approach but also an important interpretive tool for uncovering the links between spectral information and crop physiology. Unlike VIP, which evaluates variable contributions post hoc, Lasso incorporates an L1 regularization term into the training objective, which minimizes fitting error while shrinking many unimportant coefficients to zero, thereby achieving “automatic feature selection” and producing a sparse solution. This mechanism makes it particularly powerful when handling hyperspectral data where the number of variables far exceeds the number of samples (p ≫ n). The Galán group [84,85] found in predicting winter rye biomass that although Lasso and Elastic Net (EN) achieved comparable predictive performance, Lasso required only 32 bands to match the performance of EN, which needed 54 bands, thus offering a more concise “spectral fingerprint.” More innovatively, these selected bands were used to build a hyperspectral prediction model (HBLUP), whose predictive performance under complex genetic backgrounds even surpassed that of traditional genomic prediction models (GBLUP). This provides strong evidence of Lasso's great potential as an advanced strategy for information extraction and data compression.

When relationships in data extend beyond linear assumptions, tree-based ensemble algorithms demonstrate distinct advantages. Models such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost) effectively capture nonlinear relationships and complex interactions among spectral bands, while simultaneously evaluating the importance of each feature during training. In the iterative construction of base learners, feature importance is naturally estimated according to their contributions to model performance (e.g., information gain or impurity reduction). To enhance the robustness of this process, the Boruta algorithm was developed. Rather than serving as a new model, Boruta acts as a wrapper around RF, providing a stricter feature selection framework. It introduces “shadow features” (randomly permuted copies) as references in statistical tests, ensuring that selected features are significantly more important than random noise, thereby reducing spurious correlations. For example, Wijesingha et al. [86] applied Boruta to hyperspectral data to identify the most relevant bands for monsoon crop parameters (LAI, LCC, etc.), thereby mitigating overfitting in subsequent modeling; Li et al. [51] also confirmed its effectiveness in selecting sensitive bands for winter wheat yield prediction (R² = 0.78). Meanwhile, XGBoost represents another mainstream technique alongside RF. It likewise estimates feature importance through information gain during iterations and has gained wide adoption for its efficiency and accuracy. Peng et al. [31] employed XGBoost for band selection and integrated it with deep learning models, achieving a high predictive accuracy for rice leaf nitrogen content (R² = 0.89).

In practice, no single embedded feature selection method is universally optimal; its effectiveness is highly dependent on the specific application and estimation target. This was clearly demonstrated by Wang et al. [74], who found that Random Forest (RF) performed best for estimating aboveground biomass (AGB) (R² = 0.68), while Competitive Adaptive Reweighted Sampling (CARS) yielded superior results for predicting leaf nitrogen content (LNC) (R² = 0.87). This task-dependent nature has driven a shift toward integrating multiple methods to achieve more stable and robust feature selection. For instance, Sankararao et al. [36] combined seven approaches—including SFM-RF (feature selection based on RF importance)—and applied clustering and entropy-based criteria to refine the features selected by all methods, ultimately identifying five optimal bands. This approach achieved an impressive 96.46 % accuracy in early detection of water stress in peanut canopies. Similarly, Chancia et al. [87] developed an ensemble framework incorporating multiple algorithms for estimating nutrient content in grape leaves. These studies suggest a future trend in hyperspectral feature selection: shifting away from evaluating individual methods in isolation and toward the development of adaptive, hybrid strategies that integrate multiple approaches to maximize information extraction and optimize model performance.

In summary, embedded methods offer a distinctive balance of efficiency and performance in hyperspectral band selection. Whether based on linear frameworks that emphasize model interpretability and sparsity (e.g., VIP, Lasso), or nonlinear ensemble approaches that capture complex relationships (e.g., Random Forest, XGBoost), these methods integrate feature selection directly into the model training process. However, extensive empirical evidence indicates that the effectiveness of any single embedded method is highly context-dependent—there is no universally optimal solution. This task-specific nature has given rise to a key methodological trend: a shift away from identifying a singular “best” method toward developing adaptive, multi-method integration strategies and hybrid frameworks. By combining the outputs of multiple algorithms, researchers aim to achieve more stable and robust feature selection that can better accommodate the complexity and variability of agricultural environments. This approach not only enhances model performance but also signals a future direction for hyperspectral feature selection—one that emphasizes intelligent, strategic integration to maximize information acquisition and optimize predictive accuracy.

4.4. Summary of feature selection methods

In UAV-based hyperspectral crop monitoring, three major frameworks dominate the field of band selection: filter methods, wrapper methods, and embedded methods. These approaches reflect distinct priorities—efficiency, performance, and a balance of both, respectively. Filter methods, such as correlation coefficient (CC) analysis and successive projections algorithm (SPA), rely on statistical criteria to rapidly eliminate redundant features. Their key strength lies in computational efficiency, especially for high-dimensional datasets. However, because they evaluate features independently of model training, they are limited in capturing nonlinear relationships and feature interactions. Wrapper methods, in contrast, treat feature selection as an optimization problem guided directly by model performance. This allows them to model complex interactions and achieve higher predictive accuracy. Representative techniques include recursive feature elimination (RFE), multiple stepwise regression (MSR), and metaheuristic-based global search strategies (e.g., genetic algorithms and shuffled frog leaping algorithm, SFLA). These methods often yield superior feature subsets in complex scenarios but come with increased computational costs and a higher risk of overfitting, making them more suitable for fine-tuning after an initial round of filtering. Embedded methods integrate feature selection into the model training process, offering a compromise between efficiency and performance. Techniques such as variable importance in projection (VIP) and Lasso regression in partial least squares models, along with tree-based methods like Random Forest and XGBoost, strike a balance by maintaining low computational demands while considering feature interdependencies and offering some level of interpretability. To clearly illustrate the characteristics and differences among these methods, Table A1 in the Supplementary Material presents a systematic comparison of their principles, strengths, limitations, computational complexity, and suitable applications.

Overall, these three types of methods are not mutually exclusive but rather complementary, each serving a different purpose. Filter methods are well-suited for initial dimensionality reduction; wrapper methods excel at optimizing feature subsets and modeling complex relationships; and embedded methods provide a middle ground with both efficiency and effectiveness. In recent years, a hybrid “coarse-to-fine” paradigm has become mainstream—first applying filter methods for rapid dimensionality reduction, followed by wrapper or embedded methods for deeper refinement. Moreover, growing evidence suggests that the optimal feature selection strategy is not static. It must adapt dynamically to crop type, growth stage, and target trait. This trend indicates a shift in future research from comparing individual algorithms to developing adaptive, intelligent frameworks that integrate multiple methods, aiming to maximize information extraction and optimize model performance in complex and variable agricultural environments.

5. Band feature extraction methods

Feature extraction plays a critical role in HSI data processing, particularly in the analysis of crop biophysical properties. In the previous sections, we discussed feature selection methods, which focus on identifying the most representative features from existing data to reduce redundancy. In contrast, feature extraction involves transforming the original band features into a new feature space through specific transformations, creating new feature combinations while retaining as much useful information from the original data as possible [13]. The process of feature extraction is illustrated in Fig. 6. Although both methods share a similar goal—enhancing data analysis efficiency and model performance—their essence differs: feature selection involves choosing existing features, whereas feature extraction creates new ones. Based on different methodologies, feature extraction methods can be categorized into four groups: statistical-based, wavelet analysis-based methods, deep learning-based methods, and knowledge-guided approaches [10]. This chapter will primarily introduce these methods and explore their specific applications and advantages in hyperspectral data analysis. Moreover, vegetation indices have been widely used in feature extraction across numerous studies, demonstrating exceptional performance in analyzing plant biophysical parameters. Methods related to vegetation indices will be discussed in detail in Chapter 6.

Fig. 6.

Workflow of Hyperspectral Band Feature Extraction. Unlike band selection. The figure shows the process of generating new features from hyperspectral data through mathematical transformations. Algorithms such as Principal Component Analysis (PCA), Minimum Noise Fraction (MNF), Independent Component Analysis (ICA), and vegetation index construction methods (e.g., NDVI, OSAVI, NDRE) are used to map the original bands {b₁, …, b_i} into new feature sets {r₁, …, r_n}. These transformed features, calculated as f{b_i}, compress data dimensionality and enhance sensitivity to crop physiological traits, thus improving model interpretability and performance.

5.1. Statistical methods

In UAV hyperspectral imaging, statistical dimensionality reduction methods are among the most commonly used approaches for feature extraction. The primary objective of these methods is to use mathematical models to project high-dimensional data into a lower-dimensional space while preserving as much original information as possible [13]. This section will provide an in-depth discussion of statistical dimensionality reduction methods such as PCA, MNF, ICA, PLSR, and NMF, all of which are widely applied in UAV-based hyperspectral crop analysis. While these methods share strengths in efficient data compression and decorrelation, their methodological evolution reflects distinct analytical goals—ranging from uncovering intrinsic data structures, to incorporating supervised information, and ultimately to applying physical constraints.

The foundation of this paradigm is principal component analysis (PCA) [20], an unsupervised feature extraction method designed to maximize variance. By applying orthogonal transformations, PCA reshapes raw spectral data into a set of linearly uncorrelated principal components (PCs). This process enables efficient data compression and denoising while preserving the major sources of variation. In crop parameter retrieval and monitoring, the core value of PCA lies in its ability to efficiently distill key features. In practice, PCA is frequently employed as a powerful preprocessing step for a wide range of modeling approaches [88]. For instance, Mozgeris et al. [89] used PCA-derived features to estimate chlorophyll content in spring wheat. When these features were input into a geographically weighted regression (GWR) model, prediction accuracy improved significantly, achieving an R² of 0.77. Similarly, Wen et al. [80] combined PCA with Gaussian process regression (GPR) to retrieve nitrogen content in rice with high accuracy (R² = 0.85). These studies demonstrate the robustness of PCA across different crops and modeling frameworks. With the advent of deep learning, this “first extract, then model” strategy has become increasingly important. A representative example is provided by Ma et al. [44], who first reduced 75 near-infrared hyperspectral bands (750–1000 nm) into three principal components using PCA and then input them into a deep learning network, successfully predicting winter wheat yield (R² = 0.74). Moreover, PCA exhibits considerable flexibility: its application is not limited to raw spectral bands but can also serve as a versatile analytical tool. For example, Hu et al. [47] innovatively applied PCA to a set of vegetation indices (VIs), effectively addressing collinearity among indices and achieving accurate biomass estimation for Chinese milk vetch during peak flowering (R² up to 0.81). PCA is also widely used as an exploratory tool; for instance, Colovic et al. [90] employed it to reveal intrinsic relationships among data collected from different sensors. However, the fundamental limitation of PCA lies in its lack of physical interpretability: the newly generated principal components are mathematical combinations of all original bands, and thus lack explicit physical meaning.

To address the inherent limitations of PCA, a series of advanced variants have been developed. When the signal-to-noise ratio is the primary concern, minimum noise fraction (MNF) analysis offers a superior solution. Unlike PCA, which focuses solely on maximizing variance, MNF aims to maximize the signal-to-noise ratio. It follows a two-step procedure—noise whitening followed by PCA decomposition—which enables more effective separation of signal and noise in hyperspectral imagery. In practice, MNF is widely used as an efficient preprocessing tool. For example, Kanning et al. [91] applied MNF for spectral denoising prior to PLSR modeling in winter wheat yield estimation. This approach achieved an R² of 0.79 for leaf area index (LAI; RMSE = 0.18 m²/m²) and 0.77 for chlorophyll content (CHL; RMSE = 7.02 μ g/cm²). Furthermore, when the estimated LAI and CHL were used as inputs, a simple multiple linear regression model accurately predicted final yield with an R² as high as 0.88. These results demonstrate that effective data purification through MNF is a critical prerequisite for subsequent models to accurately capture crop physiological status.

When the objective is to separate signal sources rather than simply remove correlations, independent component analysis (ICA) offers a clear advantage. Unlike PCA and MNF, which rely on second-order statistics such as variance and signal-to-noise ratio, ICA employs higher-order statistics to separate statistically independent signal sources. This mechanism enables ICA to unmix hyperspectral data and isolate independent signals corresponding to specific biochemical components, which may exhibit small variance yet carry critical information [92]. In a study on tea plants, Tu et al. [93] systematically compared PCA, MNF, and ICA. Their objectives included estimating biochemical parameters such as tea polyphenols. They found that ICA, when combined with support vector machines (SVM), achieved an accuracy of 93.8 % in distinguishing different cultivars—substantially outperforming PCA and MNF. This is because PCA and MNF tend to retain components with maximum variance and SNR, which can sometimes overlook independent features critical for multi-class classification tasks. By unmixing independent signals, ICA is more effective in retaining detailed feature information required for classification and prediction tasks. These findings underscore ICA's remarkable ability to isolate subtle yet decisive independent features.

While the aforementioned unsupervised methods excel at exploring intrinsic data structures, PLSR introduces feature extraction into the supervised domain. The fundamental distinction of PLSR lies in its extraction of latent variables: it not only accounts for the variation within spectral data (X) but also seeks to maximize covariance between spectral data and target physiological or biochemical parameters (Y). Compared to traditional methods such as PCA and MNF, PLSR not only focuses on maximizing the variance of predictors but also incorporates response variable information, excelling in regression analysis. Within complex modeling pipelines, PLSR often serves as a powerful feature engineering tool, distilling key information for subsequent machine learning models [94]. For example, Sahoo et al. [95] extracted critical latent variables using PLSR and employed them as inputs for an artificial neural network (ANN). This approach achieved an R² of 0.983 for leaf area index (LAI) and an even higher R² of 0.998 for canopy chlorophyll content (CCC) in winter wheat. Ban et al. [96] highlighted the stability of PLSR in model transferability during rice chlorophyll estimation. They found that although models such as SVR achieved slightly higher accuracy in single-region validation (R² up to 0.86), their performance declined sharply under cross-regional validation. By contrast, models built on PLSR-extracted features maintained robust predictive accuracy (R² $>$ 0.7), demonstrating exceptional generalization ability. To further enhance performance, researchers have continued to refine the method. For instance, Geipel et al. [97] applied weighted partial least squares regression (WPLSR) to estimate forage quality, assigning higher weights to sensitive spectral regions. This approach yielded R² values of 0.80 for dry matter yield and 0.72 for crude protein content.

Finally, when physical interpretability is prioritized, non-negative matrix factorization (NMF) offers a unique nonlinear feature extraction approach. The key advantage of NMF lies in its non-negativity constraint, which ensures that both the basis and coefficient matrices retain physical meaning. This makes it an ideal tool for addressing mixed-pixel problems and performing spectral unmixing. The value of NMF was clearly demonstrated by Lu et al. [98], who successfully separated pure vegetation spectra from mixed rice field hyperspectral data using NMF. Subsequent modeling improved the prediction accuracy of potassium accumulation, raising the R² from 0.71 to 0.83. However, the study also noted that performance is closely linked to vegetation cover. During the early tillering stage, when cover was sparse ( < 23 %), the method's effectiveness was limited, thereby defining its application boundaries.

In summary, statistical feature extraction methods form a comprehensive technical framework that spans unsupervised to supervised and linear to nonlinear approaches. This progression begins with PCA-based unsupervised exploration of data structure and extends to specialized variants, such as MNF for maximizing signal-to-noise ratio and ICA for isolating independent signals. Supervised methods, exemplified by PLSR, incorporate target variable information, thereby tightly coupling feature extraction with regression tasks and enhancing predictive capability. Finally, nonlinear constraint-based methods, represented by NMF, provide critical solutions for challenges such as spectral unmixing and maintaining physical interpretability. Thus, selecting among these methods fundamentally involves a trade-off between predictive accuracy, computational efficiency, and the physical interpretability of extracted features. Although this paradigm is highly powerful, most of its methods sacrifice the physical meaning of original spectral bands. This limitation has fostered the development of alternative feature extraction approaches, such as treating spectra as signals for multiscale decomposition (e.g., wavelet analysis) or leveraging domain knowledge to focus directly on features with clear physical meaning. These approaches will be elaborated in subsequent sections.

5.2. Wavelet analysis methods

Unlike the statistical paradigm that treats each spectral band as an independent variable, signal processing–based techniques consider the entire spectrum as a continuous one-dimensional signal. This perspective aims to uncover deep structural features, often difficult for purely data-driven methods to capture, by applying mathematical transformations in the frequency or time–frequency domains. Wavelet analysis is the most representative and widely applied technique within this paradigm. This technique primarily includes Continuous Wavelet Transform (CWT) and Discrete Wavelet Transform (DWT), with CWT being extensively used for extracting crop parameter features from hyperspectral canopy reflectance [99]. The fundamental principle of CWT is to decompose input signals into components of different frequencies, offering scale-appropriate resolution to effectively reveal critical spectral features in the signal. This capability makes CWT an effective tool for extracting critical spectral features from hyperspectral reflectance spectra of vegetation canopies [99]. The strength of wavelet analysis in overcoming the limitations of traditional vegetation indices has been well demonstrated in applications such as biomass estimation. For example, Yue et al. [99] combined image wavelet decomposition (IWD) with CWT to accurately estimate aboveground biomass (AGB) in winter wheat. This combination effectively overcame the saturation problem observed with traditional optical indices under high canopy cover, achieving an estimation accuracy of R² = 0.85—substantially higher than models relying solely on conventional indices. However, wavelet-based features are not always the optimal choice. In a comparative study, Sun et al. [100]conducted a comprehensive comparison between wavelet coefficients extracted using continuous wavelet transform (CWT), raw spectral data, first-order derivatives, and various vegetation indices. The results showed that while wavelet features (R² = 0.61) outperformed raw spectra and derivatives, a model based on selected vegetation indices slightly outperformed them (R² = 0.63) under that specific stress condition. This finding indicates that the choice of optimal features is highly context-dependent. In fact, the greater value of wavelet analysis may lie in its role as a catalyst for integrating multi-source remote sensing information, thereby achieving accuracies unattainable with single data sources. A representative example is provided by Liu et al. [101], who creatively fused traditional VIs, high-frequency texture features from RGB images extracted using DWT, and energy coefficients from hyperspectral imagery derived via CWT. Their results confirmed that this multi-source feature fusion strategy significantly improved the estimation accuracy of potato AGB. The model achieved an R² of 0.72 on the validation set, with RMSE reduced by more than 27 % compared with any single-feature approach—clearly demonstrating the potential of wavelet analysis to synergize diverse remote sensing information.

5.3. Deep learning-based methods

Both statistical transformations and wavelet analysis rely heavily on predefined mathematical rules for feature extraction. Deep learning fundamentally changes this paradigm by constructing deep nonlinear networks, enabling a shift from “manually designed features” to “data-driven automatic representation learning.” Although deep learning techniques involve relatively high computational costs, their superior performance in handling high-dimensional data, such as hyperspectral data, has been widely validated [102]. The starting point of this paradigm is the autoencoder (AE), an unsupervised encoder–decoder architecture that compresses high-dimensional spectra into low-dimensional representations, enabling nonlinear feature extraction. To address the challenges of high similarity between adjacent hyperspectral bands and large amounts of redundant information, researchers have employed AEs to extract features from reflectance data, thereby reducing redundancy and improving modeling efficiency. In the study by Yu et al. [103], an AE was used to extract features from rice hyperspectral data to estimate the nitrogen nutrition index (NNI). By compressing the spectra into 30 key features, the AE effectively reduced redundancy. Combined with an optimized machine learning model, the approach achieved an R² of 0.837 on the validation set, significantly outperforming alternative methods. Building on the success of AEs, convolutional autoencoders (CAEs) addressed the limitation of AEs being confined to one-dimensional spectral features, further enhancing feature extraction. CAE combines the spatial information capturing capabilities of Convolutional Neural Networks (CNN) with AE's dimensionality reduction characteristics, enabling more efficient feature extraction in both spectral and spatial dimensions. In the study by Paul et al. [102], a CAE was used to extract 32 deep features from 264 spectral bands for predicting canopy-average chlorophyll content (CACC) in pear trees. When these features were input into a Gaussian process regression (GPR) model, the prediction error (RMSE) was reduced by 7.03 % compared with models using raw bands, demonstrating the superiority of CAEs in capturing joint spectral–spatial information. In addition to AE and CAE, other innovative deep learning methods have also shown potential in feature extraction. For instance, Yue et al. [104] proposed a hyperspectral-to-image transformation (HIT) method that converts one-dimensional spectra into two-dimensional images, then applies convolutional neural networks (CNNs) with transfer learning to estimate soybean leaf chlorophyll content (LCC). Their Soybean-LCCNet model achieved strong performance in independent validation (R² = 0.78), far surpassing the best-performing traditional model in the study (PLSR, R² = 0.61). This highlights the advantages of deep learning in both feature representation and model generalization.

In summary, deep learning–based methods represent a fundamental shift in feature extraction from manual design to automatic learning. From nonlinear spectral compression with AEs to joint spectral–spatial representations with CAEs and innovative architectures such as HIT + CNN, this paradigm continuously pushes the boundaries of feature representation and predictive accuracy. However, these advances come with trade-offs, including reliance on large datasets, high computational costs, and challenges of interpretability due to the “black-box” nature of the models. These factors must be carefully considered in practical applications. Thus, the deep learning paradigm can be regarded as a “high-investment, high-reward” pathway, trading off interpretability and computational resources for the potential of significant performance breakthroughs.

5.4. Knowledge-guided approaches

Unlike the purely data-driven transformation methods discussed earlier, the central idea of this set of strategies is to deeply integrate domain-specific prior knowledge into the feature extraction process, thereby generating features that are more interpretable and task-oriented. Vegetation indices represent one of the most classical examples of this approach, and they will be discussed in detail in subsequent sections. This integration of knowledge is reflected at two levels: first, at the spectral–physical level, by leveraging our understanding of the mechanisms underlying interactions between spectra and matter; and second, at the sample level, by incorporating knowledge of spatial heterogeneity in agricultural production environments.

At the spectral–physical level, spectral analysis methods achieve knowledge-driven feature extraction by amplifying or isolating spectral characteristics with clear physiological or biochemical meaning. Among the most classical techniques are first derivative (FD) spectra and continuum removal (CR), both of which enhance local variations in spectral curves and thereby highlight absorption features associated with crop biochemical components such as pigments and nitrogen. For instance, Zhu et al. [54] applied FD and CR to construct new spectral indices, improving the coefficient of determination (R²) of winter wheat leaf nitrogen content (LNC) estimation models to 0.815—an accuracy increase of 45 % compared with traditional nitrogen indices. Similarly, Wang et al. [60] used FD spectra to identify sensitive bands, achieving validation R² values of 0.832 and 0.797 for soybean leaf chlorophyll content (LCC) and nitrogen balance index (NBI) models, respectively—demonstrating the robustness of FD in multi-parameter inversion. However, their effectiveness is not universally optimal, and must often be compared against alternative feature extraction methods, with performance depending on specific contexts. This was clearly demonstrated in a study by Sun et al. [100] on monitoring maize canopy health under lodging stress, where the authors systematically compared raw spectra with multiple transformed features for estimating canopy chlorophyll density. The results showed a clear performance hierarchy: all feature extraction methods outperformed raw spectra (R² = 0.57), with optimized vegetation index (VI)-based models performing best (R² = 0.63), slightly ahead of wavelet coefficients (R² = 0.61) and FD spectra (R² = 0.59). These findings strongly suggest that under complex stress conditions, the choice of feature extraction method must be tailored to context. To address more complex sources of interference, researchers have developed targeted algorithms such as directional second derivative (DSD), which suppress soil and viewing angle effects. Using this approach, Yang et al. [105] successfully improved the R² of winter wheat yield estimation models to 0.787. Going further, spectral analysis techniques can even be applied to directly probe core physiological functions of crops, such as photosynthesis. Zarco-Tejada et al. [106] pioneered the use of fluorescence retrieval based on the Fraunhofer Line Depth (FLD) principle, successfully inverting sun-induced chlorophyll fluorescence (SIF), which is directly related to photosynthesis. The retrieved SIF showed a correlation of R² = 0.54 with ground-measured net photosynthetic rate (Pn), demonstrating the potential for direct, non-destructive monitoring of core physiological functions in crops.

Beyond integrating knowledge at the spectral–physical level, another powerful strategy is to incorporate prior understanding of sample-level heterogeneity in agricultural production environments. The central premise is that fields are rarely homogeneous; soil fertility, moisture, and other conditions often vary spatially. Building on this, clustering methods are employed to implement a “divide-and-conquer” strategy. Rather than directly transforming features, this approach first partitions a heterogeneous dataset into relatively homogeneous subsets using clustering algorithms (e.g., K-means), after which tailored local models are constructed for each subset [107]. For example, Yang et al. [108] used UAV hyperspectral data to estimate SPAD values of winter wheat at the jointing stage but faced spectral heterogeneity caused by uneven fertilizer application across the field. To address this challenge, they applied K-means clustering based on spectral angle distance to divide all spectral samples into distinct clusters, effectively separating samples associated with different soil fertility conditions. For each cluster, they independently trained an XGBoost regression model using all 176 spectral bands as input features. The results showed remarkable success: the optimal Cluster-XGBoost model achieved an R² of 0.925, representing an 18.4 % improvement in accuracy and a reduction in root mean square error (RMSE) of more than 50 % compared with the conventional single global model trained on all samples.

In summary, whether by enhancing physical features through spectral analysis or addressing field heterogeneity via spectral sample clustering, these strategies share a common foundation: the integration of domain knowledge into feature engineering. This knowledge-informed approach transcends purely data-driven methods, offering more targeted, interpretable, and robust solutions to the specific challenges of agricultural remote sensing.

5.5. Summary of feature extraction methods

Unlike feature selection, which aims to identify subsets of the original bands, the feature extraction methods discussed in this chapter focus on transforming and restructuring high-dimensional spectral data to generate a new set of features that are lower in dimension but richer in information. In UAV-based hyperspectral crop monitoring, feature extraction can be broadly categorized into three mainstream approaches: statistical transformations, signal processing and physics-based methods, and automated representation learning. These represent different strategies ranging from optimizing data structure, to enhancing physical interpretability, to achieving automated learning. Classical statistical approaches are the most widely applied, with principal component analysis (PCA), independent component analysis (ICA), and partial least squares regression (PLSR) as representative techniques. These methods compress data efficiently and remove collinearity through mathematical transformations, though their objectives differ: PCA maximizes variance, ICA seeks signal independence, and PLSR maximizes covariance with target variables under a supervised framework. However, the common drawback is that the new features lose the clear physical meaning of the original bands, thereby reducing model interpretability. The second stream is rooted in signal processing and physical mechanisms, aiming to extract features with explicit meaning. Signal processing methods, such as wavelet analysis, leverage multi-scale decomposition to effectively address non-linear issues such as signal saturation. Physics-driven methods, represented by spectral derivatives and continuum removal, directly amplify or isolate spectral signals associated with specific physiological and biochemical processes in crops. The shared advantage of these approaches lies in their strong interpretability, serving as a bridge between data and underlying mechanisms. The third stream is nonlinear representation learning via deep learning, exemplified by autoencoders (AEs/CAEs). These models can automatically learn optimal low-dimensional abstract features in an end-to-end manner, and their powerful representational capacity often translates into superior predictive accuracy. However, challenges include their “black-box” nature, reliance on large-scale datasets, and substantial computational cost—factors that must be balanced when pursuing peak performance. To systematically illustrate their characteristics, Table A2 in the Supplementary Material provides a detailed comparison of the implementation principles, key strengths and weaknesses, and best-practice scenarios of the above feature extraction methods in crop monitoring.

In summary, the choice of feature extraction methods fundamentally represents a trade-off among model performance, computational cost, and interpretability of results. No single method can adequately address all challenges, and current research shows a clear trend toward multi-source feature fusion—for example, feeding PCA-reduced features into deep networks, or combining wavelet- and texture-based features. Looking ahead, the key direction lies in exploring intelligent integration of prior physical knowledge with powerful data-driven techniques such as deep learning. Such hybrid approaches can generate a new generation of features that combine high accuracy with strong interpretability, thereby elevating UAV-based hyperspectral remote sensing to new levels of application and impact.

6. Vegetation index-based feature extraction

6.1. Strategies for screening and constructing vegetation indices

In UAV-based hyperspectral research for crop biophysical parameter inversion and monitoring, feature extraction methods based on vegetation indices hold a prominent position and are widely applied. According to electromagnetic wave theory, different substances exhibit unique and diagnostically significant spectral absorption features at specific wavelengths due to their chemical composition, molecular structure, and surface characteristics, such as water absorption features, chlorophyll absorption features, and nitrogen absorption features. These absorption features provide a foundation for the precise identification of substances. Vegetation indices combine these spectral features and use mathematical operations across multiple bands to highlight the spectral characteristics of target objects, enabling effective identification of surface categories [13].

As powerful spectral parameters, vegetation indices are generated through specific algorithms applied to one or more spectral bands. However, with hundreds of existing or potential VIs, a central challenge lies in efficiently screening or constructing indices that are most suitable for the specific monitoring task. In practice, feature selection techniques are key to addressing this challenge and generally follow two main strategies. With decades of development in remote sensing, researchers have built an extensive library of VIs, each representing the crystallization of domain knowledge for specific applications. The first strategy is screening, which involves selecting optimal features from the existing VI library. For example, Zheng et al. [39] and Wu et al. [28] successfully applied correlation analysis (CC) to identify the most relevant indices for estimating maize nitrogen concentration and assessing jujube tree health, respectively. A clear trend has emerged: VIs that combine red-edge and near-infrared bands (e.g., NDRE, MTCI) have consistently been shown to be core features strongly associated with crop physiological status. The second strategy is construction, where CC is used to guide the design of new, data-driven VIs. This approach involves systematically computing correlation maps between all possible band combinations and the target parameter. For instance, this strategy has been successfully applied to rice yield estimation [109] and potato nitrogen content prediction [26]. To improve the robustness of this approach, Ren et al. [110] proposed a “local maximum R² centroid method” to identify the most stable band combinations. Zhou et al. [111] went further by combining CC-identified sensitive bands with the 4SAIL radiative transfer model to design novel indices that reduce the influence of canopy structure [19,21]. Of course, relying solely on CC cannot capture complex feature interactions. Therefore, as discussed in earlier sections, more sophisticated wrapper- or filter-based algorithms—such as recursive feature elimination (RFE), Gaussian process regression (GPR), and ReliefF—are increasingly used to select optimal subsets of VIs, aiming to build models with superior performance. Although data-driven VIs often serve as powerful inputs for machine learning models, whether they can fully replace full-spectrum information remains an open question requiring careful evaluation. Guo et al. [112] demonstrated that although VI-based models achieved high prediction accuracy (R² = 0.90), models trained on the full spectrum performed slightly better (R² = 0.92). This suggests that full-spectrum data may still contain secondary or interaction information that advanced algorithms can exploit.

6.2. Types of vegetation indices

Benefiting from decades of accumulated domain knowledge and the aforementioned screening and construction strategies, a wide range of vegetation indices has now been developed for diverse application scenarios. Following the classification proposed by Xue et al. [113], vegetation index–based feature extraction methods can generally be grouped into three main categories. The first category comprises basic vegetation indices, such as NDVI and RVI. These indices rely on simple band combinations to reflect overall vegetation growth status. While widely used, they are often sensitive to background factors such as soil and atmospheric conditions. The second category includes optimized vegetation indices, such as EVI and SAVI. By incorporating more complex algorithms to correct background noise, these indices achieve greater robustness under variable environmental conditions, making them more suitable for precision agriculture. The third category, and the primary focus of this study, consists of physiological and biochemical indices, such as NDRE (for nitrogen estimation) and TCARI (for chlorophyll estimation). These indices are specifically designed for quantitative retrieval of physiological parameters, offering higher accuracy and supporting decision-making in precision agriculture. The following sections will provide a detailed analysis of the applications of these three categories of indices.

6.2.1. Basic vegetation indices

Among the literature reviewed in this study, the most commonly used basic vegetation index is the NDVI. NDVI, introduced by Rouse et al. [114] in 1973, is a standard tool for vegetation monitoring and is widely applied in agricultural monitoring due to its simplicity and accuracy. Its principle is based on the spectral properties of healthy vegetation, which reflects more in the near-infrared band and less in the red band. NDVI is calculated using the normalized difference between near-infrared and red band reflectance, with values ranging from −1 to 1. Higher values indicate healthier vegetation and greater coverage. In UAV hyperspectral imaging applications, Jayan Wijesingha et al. used six vegetation indices to extract features of finger millet. The results showed that NDVI (800, 670) performed well in estimating the LAI, while NDVI (750, 550) was more effective in estimating CWC [86]. Similarly, Fumin Wang et al. evaluated the performance of various vegetation indices, including NDVI, RVI, DVI, and EVI, in estimating rice yield. The results indicated that NDVI performed the best across all growth stages. For instance, combining NDVI models from different growth stages, such as NDVI (824, 728) during the booting stage and NDVI (776, 724) during the maturity stage, enabled relatively accurate rice yield estimation [115]. These studies confirm the importance of NDVI in agricultural production [116], particularly in optimizing crop management and yield prediction.

The RVI is another commonly used basic vegetation index. RVI is calculated as the ratio of NIR reflectance to red (RED) reflectance, where NIR and Red represent the reflectance in the near-infrared and red bands, respectively. The calculation of RVI is simple and intuitive, making it suitable for the rapid assessment of vegetation biomass. The principle of RVI is based on the fact that vegetation, especially healthy vegetation, typically exhibits higher reflectance in the near-infrared band than in the red band. This allows RVI to effectively characterize vegetation growth conditions and biomass. In the study by Yue et al. [117], the authors evaluated the performance of various vegetation indices in remote sensing-based estimation of AGB for winter wheat. The study tested five specific spectral bands and fourteen spectral vegetation indices, including the RVI and the NDVI. The results showed that combining crop height with spectral parameters, particularly the HR670 and HR680 bands, along with RVI and NDVI indices, significantly improved the accuracy of AGB estimation.

DVI is also one of the commonly used basic vegetation indices. DVI is calculated as the direct difference between the NIR and red (Red) bands. Although DVI is not normalized, its simplicity makes it practical for preliminary analyses, particularly for quickly assessing vegetation growth and biomass. In a study by Ma et al. [35] evaluating various vegetation indices for monitoring cotton leaf area index (LAI), a Difference Vegetation Index (DVI) constructed from the green and blue-green bands—specifically DVI(540, 525)—was identified as the most effective indicator. It exhibited the strongest correlation with LAI (correlation coefficient r = −0.7591) among all tested indices, demonstrating the great potential of DVI for accurate monitoring of cotton LAI.

In our review of the relevant literature, NDVI, RVI, and DVI were often found to be used in combination to achieve better results and more comprehensively reflect vegetation conditions. For example, Zhang et al. [118] used UAV hyperspectral data to evaluate vegetation indices related to the LNC of winter wheat. The study found that NDVI (R578, R490), DVI (R830, R778), and RVI (R490, R578), constructed using raw spectra, performed well in fitting LNC, with RVI (R490, R578) achieving the highest R² at 0. 76, followed by NDVI (R578, R490) and DVI (R830, R778) at 0. 75 and 0. 65, respectively. After variance inflation factor (VIF) screening, NDVI (R578, R490) and DVI (R830, R778) were identified as the optimal vegetation indices without multicollinearity issues. These indices, combined with texture features (e. g. , MEA490, MEA778), were used to construct an estimation model for winter wheat LNC. In a study by Xu et al. [119] that integrated spectral and texture information to improve the estimation of rice aboveground biomass (AGB), several vegetation indices—including NDVI, RVI, and DVI—were evaluated. The study highlighted a key finding: the sensitivity of AGB estimation varies across different data dimensions. In the spectral domain, vegetation indices constructed from red-edge and near-infrared bands were shown to be the most effective for estimating AGB. In contrast, within the spatial domain, texture features derived from green and red bands demonstrated superior predictive capability. Xu et al. [120] investigated 18 vegetation indices for estimating rice AGB and found that DVI, RVI, and MSR were the optimal predictors at different growth stages. Additionally, indices containing red-edge or near-infrared bands (e. g. , NDVI, DVI) performed exceptionally well in the estimations. The study also found that combining vegetation indices and texture information significantly improved estimation accuracy, with different optimal feature combinations observed across spatial resolutions and growth stages.

Additionally, the Simple Ratio (SR) is another widely used vegetation index. SR is calculated as the ratio of NIR to red band reflectance. Similar to RVI, it provides direct information about vegetation density, making it highly practical for quickly assessing vegetation health and biomass. In Zhou et al. [111]'s study, 14 vegetation indices were evaluated, including 7 flavonoid-related indices (e. g. , FI420, 710 and FL420, 690) and other structural and chlorophyll-related indices (e. g. , NDVI, RDVI, and SR760, 700). The study found that chlorophyll-normalized indices performed best, particularly the combination of FI420, 710 and SR800, 710 (CV - R² = 0. 65), which excelled in estimating LFC and mapping its three-dimensional distribution in ginkgo plantations. This is because 710 nm lies within the red-edge region and is sensitive to chlorophyll content, while 420 nm and 800 nm fall within the shortwave blue and near-infrared regions, respectively, corresponding to leaf absorption and reflectance characteristics. The combination of these three wavelengths helps to reduce the influence of canopy structure on reflectance signals and enhances the sensitivity to vegetation physiological status. In Song et al. [59]'s study, 41 vegetation indices were evaluated, revealing that the optimal vegetation index varied across different growth stages and sensors. For instance, in the late growth stages of wheat, the ASD-based SR (700, 670) demonstrated high accuracy in estimating LNC, PNC, and NNI.

The Green Index (GI), calculated based on the green light band, effectively reflects vegetation growth conditions, particularly excelling in health monitoring and biomass estimation. Yue et al. [117] significantly improved the accuracy of AGB estimation for winter wheat by combining crop height with spectral parameters such as the HR670 and HR680 bands, RVI, and NDVI. The study found that models using crop height alone achieved accuracy comparable to those using a single vegetation index. However, when height and spectral data were combined, the estimation accuracy improved significantly, providing strong support for winter wheat AGB estimation and showcasing its potential in agricultural management. In a comprehensive study conducted by Zhang et al. [68], the performance of 29 vegetation indices—including DVI, GI, GRVI, and RVI—was systematically evaluated for monitoring various physiological and biochemical parameters of winter wheat. One key conclusion of the study was that different vegetation indices exhibit significantly varying sensitivities to specific crop parameters, highlighting the necessity of a “targeted index for a targeted trait” approach. The study found that indices such as IPVI, MSAVI, TVI, and SPVI performed best in estimating leaf area index (LAI); whereas for monitoring chlorophyll content (SPAD) and aboveground biomass (AGB), simpler indices such as DVI, GI, GRVI, and RVI demonstrated better suitability.

Other basic vegetation indices include the Red-Green Index (RGI) [121], Ratio Spectral Index (RSI), Normalized Difference Index (NDI) [122], Normalized Green-Red Difference Index (NGRDI) [123], Excess Green Index (ExG) [124], and Excess Red Index (ExR) [62]. For example, RGI is calculated as the ratio of red to green band reflectance and is effective in distinguishing different land cover types. RSI is commonly used to monitor vegetation health and its response to environmental changes. NDI is a flexible normalized difference form applicable to any two bands, reducing inter-band influence and enabling accurate vegetation condition assessments. NGRDI uses the normalized difference between green and red bands to monitor vegetation growth status, while ExG emphasizes the difference of green light relative to red and blue light, making it suitable for quickly identifying vegetated and non-vegetated areas. Although these indices are less frequently used than NDVI, RVI, and DVI. They still offer valuable tools and methods for vegetation monitoring and analysis. The Basic vegetation indices discussed above can be found in Table A3 in the Supplementary Material.

6.2.2. Vegetation indices for mitigating influencing factors

In UAV-based hyperspectral remote sensing, soil background and atmospheric effects are two critical factors influencing the accuracy of crop physiological and biochemical parameter inversion and monitoring. First, soil background can significantly interfere with vegetation reflectance signals, particularly in areas with low vegetation coverage. The complex interactions between soil spectral properties (e. g. , brightness and color) and crop reflectance can obscure the true signals from crops. This issue is especially prominent in arid environments or during the early growth stages of crops. Additionally, variations in soil moisture content and type can alter spectral responses, further complicating data interpretation [113,125,126]. Therefore, addressing and mitigating the soil background effects is crucial for accurate inversion and monitoring of crop physiological and biochemical parameters. Aerosols, clouds, and water vapor in the atmosphere cause scattering and absorption of light signals, introducing noise and reducing the effectiveness of the data. These interferences are particularly pronounced under adverse atmospheric conditions, such as high humidity or air pollution, affecting the accurate inversion and monitoring of crop physiological and biochemical parameters. To address these challenges, vegetation indices designed to eliminate soil background and atmospheric effects have been developed [113,125,126]. These indices are designed with specific algorithms to minimize the impact of environmental factors on inversion and monitoring outcomes, thereby improving the accuracy and reliability of remote sensing data. Details of the vegetation indices for eliminating influencing factors are provided in Table A4–A6 of the Supplementary Material.

Mitigating Soil Interference: A critical challenge in UAV-based hyperspectral remote sensing is minimizing the influence of soil background on accurately retrieving and monitoring crop physiological and biochemical parameters. To address this, various vegetation indices have been developed, among which the Optimized Soil-Adjusted Vegetation Index (OSAVI) is one of the most widely used. OSAVI incorporates reflectance from the NIR and RED spectral bands and employs a soil adjustment factor to mitigate soil interference in low-density vegetation signals while maintaining sensitivity to high-density vegetation. For instance, Zhu et al. [127] analyzed OSAVI's performance in estimating LAI and the chlorophyll content (Cab) using the PROSAIL model and proposed a dual-layer matrix method to reduce cross-interference between LAI and Cab. Similarly, Domingues et al. [128] evaluated 21 vegetation indices and concluded that OSAVI was most effective in estimating ground cover for organic potatoes. Additionally, the Soil-Adjusted Vegetation Index (SAVI) and its derivatives, such as MSAVI and MSAVI2, are extensively applied to reduce soil background effects, especially in sparsely vegetated areas. SAVI enhances estimation accuracy under varying soil conditions by introducing a soil adjustment factor to the NDVI framework. Cilia et al. [129] used hyperspectral remote sensing data to assess the nitrogen content in maize fields, demonstrating that MCARI/MTVI2 and MTVI2 performed exceptionally well in estimating leaf nitrogen concentration and biomass, respectively. Other indices designed to correct for soil background effects include the Differential Spectral Index (DSI) and the Modified Normalized Difference Index (MND). DSI is particularly suited for change detection in complex vegetation environments, while MND improves index accuracy through correction terms [51,55]. By leveraging distinct methodologies, these indices effectively reduce the soil background interference, significantly enhancing the accuracy of crop physiological and biochemical parameter retrieval and monitoring across diverse applications.

Mitigating Atmospheric Interference: Variations in atmospheric conditions often degrade the accuracy of remote sensing image analysis. To address this challenge, researchers have developed various vegetation indices to reduce the impact of atmospheric scattering and absorption on remote sensing data. Key examples include the Visible Atmospherically Resistant Index (VARI) and the ARVI. VARI, designed using visible light bands (red, green, and blue), effectively minimizes atmospheric effects and performs well under significant atmospheric variation. Sankararao et al. [48] demonstrated the efficacy of VARI in optimizing atmospheric correction for UAV-based hyperspectral data, significantly improving the detection accuracy of early water stress in peanut vegetation. Similarly, Liu et al. [33] employed VARI to enhance the estimation accuracy of the NNI for winter rapeseed, confirming its effectiveness in multispectral image analysis. ARVI, developed by Kaufman and Tanre, introduces the blue band to the NDVI framework to compensate for atmospheric aerosol effects, thereby improving its resistance to atmospheric interference. Wang et al. [74] utilized ARVI to mitigate atmospheric scattering effects in vegetation monitoring, achieving higher accuracy in monitoring winter wheat growth. By integrating hyperspectral and RGB imagery, their study effectively estimated AGB and LNC in wheat, providing valuable insights for agricultural management and fertilizer optimization.

Integrating Soil and Atmospheric Interference: Some vegetation indices are designed to simultaneously address the effects of both soil and atmospheric interference, significantly improving the interpretability of remote sensing data. Among these, the EVI is specifically developed to optimize vegetation signal extraction and enhance monitoring accuracy. EVI integrates reflectance from NIR, RED, and blue (BLUE) bands, using specific correction coefficients to reduce atmospheric scattering (particularly aerosols) and soil background effects. Compared to traditional NDVI, EVI not only amplifies vegetation signals but also compensates for soil and aerosol influences through the blue band, making it widely applicable in global vegetation monitoring, precision agriculture, and environmental change studies. For instance, Zhang et al. [2] evaluated 30 vegetation indices to estimate maize AGB at various growth stages, finding that EVI and VARI performed best during the R1 growth stage, providing critical insights for accurate AGB estimation. Similarly, Guo et al. [130] demonstrated EVI's superior performance in retrieving LAI for maize, particularly under specific conditions, where it reliably reflected vegetation growth status. EVI2, a simplified version of EVI, is tailored for remote sensing systems lacking a blue band. While excluding the blue band, EVI2 retains EVI's core advantages and effectively reduces soil and atmospheric interference. Its simplified calculation process makes it suitable for applications with limited resources or restricted equipment. Yue et al. [99] compared EVI2 and EVI in estimating AGB for winter wheat, showing that EVI2 provided consistent predictions under equipment constraints. Furthermore, Li et al. [131] assessed EVI2's performance in AGB prediction across different growth stages, reporting exceptional accuracy in validation datasets, particularly when applied in the CBA-Wheatzs model.

6.2.3. Vegetation physiological and biochemical indices

Nitrogen-Related VIs: In UAV-based hyperspectral research for crop parameter inversion and monitoring, detecting nitrogen content is a critical task. To achieve this, many vegetation indices have been specifically designed for nitrogen estimation. First, the Normalized Difference Red Edge (NDRE) index is one of the most widely used indices for nitrogen detection. It utilizes the reflectance characteristics of the red-edge band to assess nitrogen content in plants. The red-edge band is highly sensitive to physiological changes within plants, particularly those related to nitrogen, making NDRE highly effective in agricultural applications [132]. To further improve accuracy, NDRE1 and NDRE2 were developed as optimized versions of NDRE. These indices effectively reduce interference from complex soil backgrounds, enabling more precise nitrogen estimation. Another commonly used vegetation index is the Modified Simple Ratio 705 (mSR705). Based on spectral variations at the 705 nm wavelength, mSR705 sensitively reflects nitrogen absorption by plants, making it a prevalent tool for nitrogen detection in agriculture[99,[133], [134], [135], [136]]. The Double Peak Canopy Nitrogen Index (DCNI) analyzes the reflectance characteristics of double-peak bands in the canopy, effectively distinguishing different nitrogen levels, particularly in dense vegetation cover scenarios [31,54,63,129,137,138]. The Nitrogen Reflectance Index (NRI) is specifically designed to evaluate nitrogen levels in plants. By analyzing spectral reflectance differences, NRI accurately reflects crop nitrogen content and is widely used in crop management and fertilization optimization [[139], [140], [141]]. Similarly, the Blue Nitrogen Index (BNI) detects nitrogen content through blue band reflectance. Since the blue band is particularly sensitive to chlorophyll absorption, BNI leverages nitrogen-related variations in blue reflectance, making it an effective tool for nitrogen detection [39,95]. Additionally, the Optimized Normalized Difference Index (NDIopt) is an optimized version of NDVI specifically designed for nitrogen detection. By optimizing the normalized difference of spectral bands, NDIopt achieves more accurate nitrogen estimation, particularly under complex environmental conditions. The Normalized Difference Nitrogen Index (NDNI) is specifically designed for nitrogen detection. By utilizing band reflectance differences, NDNI directly assesses nitrogen levels and is widely applied in nitrogen management [142]. The Red Edge Position (REPLE) index evaluates nitrogen content and plant health by analyzing changes in the red-edge band position. As the red-edge band is highly sensitive to plant physiological states, particularly nitrogen variations, REPLE is commonly used for monitoring plant nutritional status. The Modified Red Edge Inflection Point (mREIP) is an adjusted version of the red-edge inflection point index, focusing on nitrogen content and plant health detection. By refining the inflection point position in the red-edge band, mREIP improves detection accuracy [143]. More complex indices, such as the Double-Peak Canopy Nitrogen Index (DPCNI) and the Vogelmann Index (VOGI), are also widely applied for nitrogen detection [118,144]. DPCNI utilizes a double-peak band structure, performing well under complex canopy conditions, while VOGI integrates multiple spectral band characteristics to effectively evaluate nitrogen levels in plants. Some researchers have proposed indices optimized from traditional vegetation indices. Xiao et al. [145] proposed three improved vegetation indices:Derivative Spectral Optimized Soil-Adjusted Vegetation Index (FDOSAVI), Derivative Spectral Difference Vegetation Index (FDDVI), and Derivative Spectral Normalized Difference Vegetation Index (FDNDVI). These indices are based on traditional vegetation indices (e. g. , OSAVI, DVI, NDVI) and optimized by incorporating derivative spectral processing methods. They aim to enhance sensitivity to crop nitrogen content and more accurately reflect vegetation physiological states.

Additionally, some combinations of vegetation indices have been specifically designed to reduce soil background interference and enhance the accuracy of nitrogen detection. For instance, the MCARI/MTVI2 combines the strengths of MCARI and MTVI2, significantly reducing soil background interference in nitrogen detection. MCARI corrects for chlorophyll absorption and reflectance characteristics, making it more sensitive to changes in nitrogen content. MTVI2, on the other hand, minimizes soil effects through an optimized triangular vegetation index. Together, they provide more precise reflections of crop nitrogen content in complex environments [33,51,129,146]. Another commonly used combination is TCARI/OSAVI. TCARI enhances sensitivity to chlorophyll detection, while OSAVI improves vegetation estimation accuracy by optimizing soil reflectance adjustments. Combined, these indices can detect both chlorophyll and nitrogen content, performing exceptionally well under conditions with complex soil backgrounds. This combination is particularly suited for applications requiring simultaneous monitoring of nitrogen and chlorophyll content. The upgraded TCARI2/OSAVI2 combination, with more refined band selection and adjustments, further enhances detection performance under complex soil conditions [51,68,129,143]. Similarly, TCARI/OSAVI_RE incorporates the red-edge band into the traditional TCARI/OSAVI combination to provide more precise nitrogen content evaluation. Because the red-edge band is highly sensitive to changes in plant nitrogen levels, this combination performs exceptionally well under complex soil background conditions [2,99,133]. Table A7 in the Supplementary Material shows the vegetation indices for nitrogen detection.

Pigment-Related VIs: In plant chlorophyll detection, the Red-Edge Chlorophyll Index (CIre) is one of the most commonly used indices. It leverages the high sensitivity of the red-edge band to chlorophyll variations, effectively assessing photosynthesis and plant health. Similarly, the Green Chlorophyll Index (CIgreen) evaluates chlorophyll content through green band reflectance, making it particularly suitable for monitoring vegetation with active photosynthesis [147]. To minimize soil background interference, the Transformed Chlorophyll Absorption in Reflectance Index (TCARI) provides more precise chlorophyll detection, particularly under complex background conditions. Similarly, the Modified Chlorophyll Absorption Ratio Index (MCARI) improves detection accuracy by correcting chlorophyll absorption band characteristics, with its advanced versions, MCARI2 and MCARI3, extending its application range[[148], [149], [150]]. For large-scale remote sensing applications, the MERIS Terrestrial Chlorophyll Index (MTCI) is widely used for assessing chlorophyll content in terrestrial vegetation, performing exceptionally well in satellite-based monitoring [151]. On the other hand, the Normalized Pigment Chlorophyll Index (NPCI) assesses changes in chlorophyll and other pigments using normalization, making it suitable for analyzing different plant growth stages [152]. The Plant Senescence Reflectance Index (PSRI) primarily evaluates changes in chlorophyll and carotenoids during plant senescence, serving as an important indicator of plant health [31,51,95,144]. For quantifying low chlorophyll content in leaves, the Modified Normalized Difference 705 (mND705) is a highly sensitive tool, ideal for monitoring early photosynthesis [2,54,138,139,143]. The Leaf Chlorophyll Index (LCI) [122] is a simple and user-friendly tool for chlorophyll detection, commonly used in monitoring the health of agricultural crops. The Red Edge Position (REP) is another commonly used index, detecting chlorophyll content based on changes in the red-edge band position. It is particularly suitable for assessing plant health and nitrogen levels [95,127,142,146]. Additionally, the Structure-Intensive Pigment Index (SIPI) focuses on detecting photosynthetic pigments, while the CI offers a simple and direct method for chlorophyll detection. The Simple Ratio Pigment Index (SRPI) evaluates carotenoids and chlorophyll through simple band ratios, while the Pigment Specific Simple Ratio (PSSR) detects specific pigment contents [2,31,51,95]. Among normalized difference indices, the Normalized Difference Chlorophyll Index (NDCI) is frequently used for chlorophyll detection, while PSSRa/b/c are applied to assess various types of chlorophyll and carotenoids [54,95,142]. The Triangular Chlorophyll Index (TCI) estimates chlorophyll content using a triangular geometric spectral structure [51,68,129], while the Pigment Specific Normalized Difference (PSND) is used for broader pigment detection [31,144,153]. Moreover, the Canopy Chlorophyll Content Index (CCCI) focuses on detecting CCC, particularly suitable for large-scale agricultural monitoring [143]. The Blue-Green Pigment Index (BGI) combines blue and green bands to detect photosynthetic pigments [154]. Additional pigment-related vegetation indices are provided in Table A8 in the Supplementary Material.

Biomass-Related VIs: In addition to nitrogen content and chlorophyll estimation, vegetation indices for biomass evaluation are highly diverse. The Green Normalized Difference Vegetation Index (GNDVI) is one of the most commonly used indices. It is based on the reflectance difference between the green and NIR bands, allowing for sensitive detection of photosynthetic activity and biomass accumulation. Due to the green band's high sensitivity to photosynthesis, GNDVI provides more accurate biomass estimation and health assessments than NDVI, making it widely applicable in crop monitoring, farmland management, and ecosystem evaluation. Zhang et al. [2] demonstrated that GNDVI performed well in estimating maize biomass, with sensitivity varying across growth stages, highlighting its effectiveness at different plant development stages. Similarly, the Green Ratio Vegetation Index (GRVI) captures biomass and plant health by using the ratio of green to NIR bands. GRVI's simple ratio calculation makes it suitable for real-time monitoring of crop growth and health, particularly excelling in rapidly changing agricultural scenarios [68,96,134,145]. The Spectral-Polygon Vegetation Index (SPVI) uses multiple spectral bands to construct a multidimensional reflectance space for evaluating vegetation coverage and biomass, making it particularly suitable for large-scale vegetation monitoring [155]. For high coverage areas, the Wide Dynamic Range Vegetation Index (WDRVI) is another commonly used index. By adjusting the ratio between red and NIR bands, WDRVI enhances sensitivity in dense vegetation areas, enabling more accurate biomass estimation. It is particularly suitable for monitoring crop growth [25,117,134,153]. The Ratio of Enhanced Infrared and Red Edge (REIP) combines the red-edge and enhanced infrared bands to evaluate biomass and physiological status using their ratio. As the red-edge band is highly sensitive to chlorophyll content and plant health changes, REIP is ideal for monitoring photosynthetic efficiency, biomass accumulation, and long-term ecological studies [71,86,90,156]. The NIR/GREEN (NIR to green ratio) indices are widely used to analyze the ratio between NIR and green bands, capturing photosynthetic efficiency and biomass variations. These indices are extensively applied in productivity and health monitoring during rapid growth periods or in dense vegetation areas, particularly in agricultural decision-making scenarios [156,157]. Composite indices, such as TACRI/OSAVI, combine chlorophyll absorption and soil adjustment advantages to reduce soil background interference, enabling more accurate assessments of biomass and productivity, particularly in areas with complex soil reflectance. This combination shows strong potential for applications in precision agriculture and crop management [134]. Additionally, some researchers have proposed optimized versions of vegetation indices, such as Opt-NDVI (Optimized Normalized Difference Vegetation Index) and Opt-RVI (Optimized Ratio Vegetation Index). Using “lambda-by-lambda” band optimization algorithms, these indices extract sensitive bands from UAV hyperspectral images and optimize traditional NDVI and RVI, particularly excelling in addressing spectral saturation issues under high biomass and nitrogen conditions. These optimized indices improve biomass estimation accuracy across multiple growth stages, showing superior performance in red-edge bands (734 nm and 742 nm) [141]. Biomass-related vegetation indices are presented in Table A9 in the Supplementary Material.

Vegetation Structure, Health Status, and Stress VIs: Monitoring vegetation coverage, density, health status, and stress involves various vegetation indices. By combining different spectral bands and algorithms, these indices help assess vegetation coverage, LAI, health status, and even detect early stress signals. First, the Triangular Vegetation Index (TVI) is a commonly used index for assessing vegetation coverage. It uses the triangular geometric shape formed by spectral reflectance to capture vegetation growth status and is often closely related to LAI [158]. Based on TVI, the Modified Triangular Vegetation Index (MTVI) enhances sensitivity to high-density vegetation and health conditions, making it suitable for complex growth environments. MTVI2 further optimizes this approach, providing more precise evaluations of vegetation health, particularly excelling in detecting growth changes associated with LAI [149]. The Red-Edge Normalized Difference Vegetation Index (RENDVI) is a red-edge band-based index specifically designed for monitoring vegetation health. The red-edge band is highly sensitive to photosynthesis and chlorophyll changes, making RENDVI excellent for assessing plant physiological status, health changes, and LAI [31,47,74,96]. Similar indices, such as the Red-Edge Reflectance Scaling Index (RARS) and the Red-Edge Normalized Difference Vegetation Index (RNDVI), can capture early signals of vegetation health and LAI changes, making them particularly useful for detecting early growth stages or disruptions in photosynthesis [31,132,144,153]. The Modified Green Ratio Vegetation Index (MGRVI) uses variations in green band ratios to evaluate vegetation coverage, health status, and LAI. Compared to traditional green band indices, MGRVI provides more accurate assessments of health status and LAI, particularly in monitoring dense vegetation areas. The Angle-Insensitive Vegetation Index (AIVI) reduces interference caused by variations in solar illumination angles. It is used for monitoring plant health, coverage, and growth, offering more reliable results in remote sensing analyses under varying light conditions (e. g. , different times or weather conditions) [159].

For stress monitoring, the Red-Edge Vegetation Stress Index (RVSI) is specifically designed to detect plant stress. It can capture spectral changes during early stress stages (e. g. , water deficiency or nutrient deficiency) and indirectly reflect a reduction in leaf area [9,51,118,153]. The Plant Water Index (PWI) focuses on detecting water content within plants, making it particularly effective for identifying early drought and water stress. It provides early warning signals before visible signs of stress appear [51]. In Feng et al. ’s study, 13 vegetation indices were evaluated for monitoring winter wheat growth, leading to the development of a new Composite Growth Index (CGI). These indices included red-edge reflectance, Excess Red Index (EXR), Vegetation Atmosphere Resistance Index (VARI), and Modified Green-Red Vegetation Index (MGRVI). The study found that these indices exhibited stronger correlations with CGI compared to individual growth monitoring indices. Among these, the Linear Combination Index (LCI) emerged as the optimal spectral index across four growth stages [160]. Vegetation structure, health status, and stress-related vegetation indices are presented in Table A10 in the Supplementary Material.

7. Challenges and future directions

7.1. System-level bottlenecks and algorithm-specific challenges

7.1.1. System-level bottlenecks

Before delving into the specific limitations of various feature selection and extraction methods, it is essential to recognize two persistent system-level bottlenecks in this field: the lack of standardized workflows and insufficient cross-domain generalizability of models. These issues underlie many of the technical challenges—particularly the limited generalizability of models—and collectively hinder the reproducibility and scalability of UAV-based hyperspectral applications in agricultural contexts.

In the field of UAV-based hyperspectral remote sensing, there is a widespread lack of standardized protocols or Standard Operating Procedures (SOPs) that cover the entire workflow from data acquisition to analysis. This absence is a major barrier to data comparability and methodological reproducibility, making it difficult to conduct cross-study comparisons [161]. This bottleneck affects every stage of the process, from data collection to ground validation. During data acquisition, no widely accepted standards exist for UAV flight parameters, sensor configurations, or calibration methods. For instance, push-broom sensors are highly sensitive to platform attitude fluctuations, often introducing distortions and misalignments. Similarly, the widely used Empirical Line Calibration (ELC) method depends heavily on stable field conditions, which restricts both its general applicability and accuracy. In the preprocessing stage, radiometric, atmospheric, and geometric corrections can be performed using multiple algorithmic approaches, and differences among software implementations further exacerbate data inconsistency [[162], [163]]. Finally, in ground-truth measurement, methodological heterogeneity—ranging from high-precision destructive chemical analyses to proxy indicators obtained by portable devices—combined with mismatches in field sampling strategies and spatiotemporal scales, introduces substantial uncertainty in linking remote sensing signals with ground information.

The lack of standardization across the entire workflow represents a fundamental challenge to both methodological comparison and model generalization in this field [164]. Consequently, it further exacerbates the even more pressing issue of poor model transferability. Even with rigorously processed data, models often fail when applied to different regions, crops, or growth stages—an issue in machine learning commonly referred to as the problem of domain adaptation. The root of this difficulty lies in the inherent multidimensional heterogeneity of agricultural ecosystems. First, temporal heterogeneity arises because crop spectral properties change dramatically across phenological stages. From seedling and jointing to maturity and senescence, canopy structure (e.g., leaf angle) and biochemical composition (e.g., chlorophyll, anthocyanins) shift substantially, making it difficult for models built on a single time point to remain valid throughout the growth cycle. Second, spatial heterogeneity results from uneven distributions of soil properties and water availability within fields, making spectral divergence of the same crop under different conditions (i.e., same object, different spectra) highly prevalent. Finally, biological and environmental heterogeneity further amplifies the uncertainty of spectral signals. Different genotypes exhibit distinct spectral features due to inherent differences in leaf morphology, cellular structure, and pigment synthesis efficiency. External field management practices (e.g., irrigation and fertilization) create coupled effects on water and nitrogen status that are difficult to disentangle. Meanwhile, environmental backgrounds such as weeds, bare soil, and crop residues introduce additional confounding variables into the target spectra. Under the “high-dimensional, small-sample” constraint, these sources of heterogeneity readily induce model overfitting. Instead of capturing universal physical or causal relationships, models tend to memorize local and spurious correlations present in the training data, ultimately causing predictive performance to deteriorate sharply when applied to new scenarios [165].

7.1.2. Algorithm-specific challenges

Beyond the broader system-level bottlenecks, specific algorithms for feature selection and extraction also face inherent technical challenges.

The core challenge of band selection method: At present, most band selection algorithms exhibit limited generalization capacity. They are highly dependent on specific datasets, leading to poor performance when faced with new or slightly different datasets. When data is insufficient, these algorithms are prone to overfitting, further compromising their generalization ability. Generalization can be divided into sample-specific and task-specific generalization. The former requires algorithms to produce consistent selection results across different samples, while the latter demands effective band selection across similar but distinct tasks, such as plant chlorophyll inversion and other parameter estimation tasks. However, most existing band selection algorithms are developed for specific tasks, making them less adaptable to variations in application scenarios and data types. Secondly, there is no unified standard for determining the optimal number of bands. Determining the appropriate number of bands is a critical challenge in hyperspectral band selection. Too few bands may result in information loss, while too many bands can introduce redundancy and increase computational complexity. A common approach is to predefine the number of bands, NNN, prior to selection. However, this lacks theoretical justification and primarily relies on empirical judgment. To address this, some researchers have proposed progressive band determination methods, dynamically adjusting the number of bands to gradually optimize the selection process. Overall, fewer bands may offer greater distinctiveness but risk missing critical features, whereas more bands can lead to significant redundancy. Striking a balance between performance, computation time, and complexity remains a major challenge. In addition, performance evaluation methods remain relatively limited. Band selection performance is typically evaluated based on the predictive accuracy of crop physiological and biochemical information. However, the nature of the model and its parameter configuration can also influence the results. In hyperspectral remote sensing, the primary objectives of band selection are to reduce redundancy, retain spectral information relevant to crop physiological and biochemical traits (e. g. , nitrogen content, chlorophyll), and lower computational costs. Therefore, future research on evaluating band selection performance should not rely solely on the predictive accuracy of crop physiological and biochemical information.

The inherent limitations of feature extraction methods: First, in preserving the originality of the signal, converting the original feature subset into a new feature subset during the extraction process can result in data loss. Moreover, inappropriate transformation criteria can strip the extracted features of their intuitive physical meaning, further complicating result interpretation. Additionally, most feature extraction algorithms involve multiple parameter settings, and the selection of these parameters significantly affects algorithm performance. This makes the tuning process time-consuming, complex, and technically demanding, representing a major challenge in the application of feature extraction methods. In particular, vegetation indices, though widely used, also have inherent limitations. First, spectral mixing may result in similar reflectance characteristics between different plants and surface objects, reducing accuracy. Second, environmental factors such as climate, soil type, and light intensity can influence vegetation spectral characteristics, causing the same vegetation to exhibit different vegetation index values under varying conditions, thereby increasing the complexity of feature extraction. Although some vegetation indices aim to mitigate the effects of these environmental factors and improve feature extraction, issues such as the ambiguity of biophysical parameters and information loss persist. Additionally, the nonlinear relationships between vegetation indices and biophysical parameters, as well as variations in applicability to specific plant types, further constrain their use in feature extraction.

The universal challenge of spectral spatial information fusion: Insufficient utilization of spatial information is a significant limitation of UAV-based hyperspectral remote sensing in monitoring crop physiological and biochemical parameters. Although hyperspectral images provide abundant spectral and spatial information, most existing feature extraction and selection methods primarily focus on spectral data, with relatively little attention paid to spatial information. Only a few studies have attempted to integrate the two. Over-reliance on spectral information may overlook the critical role of spatial data in revealing specific crop parameters (e. g. , nitrogen content, chlorophyll content), resulting in extracted features that fail to adequately represent the actual physical state of crops. Moreover, there is still a lack of optimal solutions for efficiently integrating spatial and spectral information. Existing techniques are often computationally intensive and exhibit inefficiency and poor adaptability when handling large-scale datasets.

7.2. Opportunities and future directions from emerging methods

In response to the lack of standardization, the limited generalization of models, and the constraints of traditional feature engineering, emerging deep learning–based approaches are driving crop physiological and biochemical feature analysis into a new paradigm, offering unprecedented opportunities to overcome these challenges.

Enhancing Model Generalization through Data Standardization and Robust Algorithms: To systematically address the dual bottlenecks of lacking standardization and poor model transferability, future research must advance along two dimensions—data and benchmarks on one side, and algorithms and models on the other—so as to improve both feature reliability and generalization capacity. At the data and benchmark level, it is urgent to establish publicly available reference datasets enriched with detailed metadata (covering sensor parameters, flight logs, environmental conditions, and ground sampling information, etc.). Such datasets would provide a fair and reproducible platform for evaluating different feature extraction and modeling algorithms, serving as critical infrastructure for advancing the field. At the same time, more robust radiometric calibration methods should be promoted, such as real-time correction using onboard spectroradiometers, to reduce dependence on variable ambient light [162]. A key pathway is the integration of radiative transfer models (RTMs, e.g., PROSAIL) into machine learning pipelines. By embedding physical priors, RTMs guide models to learn features consistent with vegetation optical principles rather than spurious correlations from specific datasets. This integration is essential for improving robustness against sensor differences, illumination variability, and soil background effects [166]. At the algorithmic and modeling level, building on more reliable data foundations enables advanced methods to reach their full potential, shifting models from learning “data-dependent” features to capturing “domain-invariant” representations. Specifically, transfer learning (TL) leverages pretrained models to achieve efficient fine-tuning with only limited data. Self-supervised learning (SSL) can autonomously extract general spectral representations from large volumes of unlabeled data, greatly reducing reliance on costly ground sampling. Domain adaptation (DA) enforces models to learn cross-domain invariant features, thereby enhancing their applicability across new regions, time periods, or crop varieties. In addition, generative models such as Generative Adversarial Networks (GANs) can contribute by augmenting and harmonizing data, thereby directly mitigating the effects of non-standardization at the data level. In summary, creating a standardized data environment to ensure reliable features, combined with developing advanced algorithms capable of learning generalizable representations, is essential for the accurate inversion of crop physiological and biochemical information.

Deep Learning–Driven Feature Engineering: To address the limitations of traditional feature engineering—such as restricted generalization, uncertainty in determining the optimal number of features, lack of unified evaluation standards, information loss, and limited physical interpretability—deep learning is driving a paradigm shift from “manual design” to “automated learning with intelligent interpretability.” At the core of this transformation is the realization of automated feature learning. Traditional approaches rely on manually designed features (e.g., vegetation indices), a process that is subjective, time-consuming, and often unable to capture complex nonlinear relationships. In contrast, unsupervised models such as Autoencoders (AEs) leverage their classic encoder–decoder architecture to compress high-dimensional hyperspectral data into a low-dimensional latent space and then reconstruct it with high fidelity [103]. This enables the model to autonomously learn compact, essential representations of the data, thereby addressing information loss and noise issues inherent in traditional feature extraction. Variants such as Convolutional Autoencoders (CAEs) and Graph Autoencoders (GAEs) further enhance robustness by integrating spatial context from neighboring pixels during dimensionality reduction [102]. A more profound transformation lies in the emergence of Explainable AI (XAI), which is turning the “black-box” nature of deep learning into an opportunity for intelligent feature selection. On one hand, attention mechanisms allow models to dynamically assign weights, enabling them to automatically focus on the most informative spectral bands. This implicitly and adaptively addresses the challenge of determining the “optimal number of bands,” while also enhancing the generalization and task adaptability of band selection. On the other hand, visualization techniques such as Class Activation Maps (CAMs) reveal the specific spectral regions that models rely on during decision-making. This not only enhances the physical interpretability of transformed features but also provides a foundation for developing more comprehensive and objective frameworks for evaluating band-selection performance.

Achieving Deep Spectral–Spatial Integration: To address the insufficient use of spatial information, deep learning provides powerful tools for fundamentally integrating multidimensional data. Among these, three-dimensional convolutional neural networks (3D-CNNs) and graph-based methods represent the most promising and practical solutions. On one hand, 3D-CNNs directly process hyperspectral data cubes (H × W × λ) through three-dimensional convolutional kernels, enabling simultaneous deep integration of spectral and spatial features. This joint learning mechanism has proven critical in complex agricultural scenarios. For example, in the challenging task of distinguishing rice from barnyard grass, early studies demonstrated the feasibility of using hyperspectral imaging to differentiate rice, barnyard grass, and weedy rice, thereby laying the foundation for the development of more advanced recognition models [167]. More recently, Zhang et al. [168] provided compelling evidence by developing the SPA-3DCNN model, which integrates band selection. In barnyard grass recognition, their model achieved an F1-score of 0.8936, substantially outperforming Support Vector Machines (SVMs), Random Forests (RFs), and one-dimensional CNNs (1D-CNNs). The study attributed the success of 3D-CNNs to their combined strengths in hierarchical feature extraction, handling mixed pixels and boundaries, and noise suppression, underscoring their effectiveness as an architecture for enhancing monitoring accuracy in complex field environments. On the other hand, graph-based methods provide a new paradigm for analyzing irregular objects in agricultural contexts. Unlike traditional CNNs, which are constrained by fixed rectangular receptive fields, these methods transform image data into flexible graph structures by defining pixels or superpixels as nodes and their adjacency as edges. Among these, Graph Neural Networks (GNNs) are particularly powerful, as they update central nodes by aggregating features from neighboring nodes. This enables the learning of context-dependent, robust features, supporting precise analysis at the “object level” (e.g., individual plants) rather than merely the pixel level. Such object-level analysis highlights the core value of high-resolution data. As noted by Shen et al. [169], UAV-based centimeter-level resolution effectively avoids the accuracy loss caused by mixed pixels in satellite remote sensing, thus offering an invaluable foundation for fine-scale monitoring. Therefore, a key future direction lies in integrating advanced models such as GNNs with high-quality UAV data to overcome current bottlenecks. In addition, several innovative network architectures—such as the hyperspectral-to-image transformation (HIT) method, which converts spectra into images for processing with CNNs [104]—have also emerged, together constituting core technical pathways for addressing the challenge of spectral–spatial information fusion.

7.3. Economic feasibility, scalability, and challenges in real-world applications

Although UAV-based hyperspectral technology has demonstrated great potential in research, its adoption and integration into real-world agricultural production systems must overcome the significant gap between technical feasibility and economic and operational feasibility. At present, high deployment costs and stringent requirements for operator expertise are among the core barriers preventing widespread adoption [161]. Specifically, the primary economic barrier lies in the high upfront investment. A professional-grade UAV hyperspectral system may cost tens to hundreds of thousands of dollars, creating a capital threshold that most farmers cannot afford. Alongside this are technical and human resource barriers. The full data-processing workflow is highly complex and requires expertise across multiple disciplines such as remote sensing and agronomy, making it inaccessible to smallholder farmers. In addition, UAV operations are vulnerable to climatic conditions, and in many rural areas the absence of reliable internet connectivity, charging stations, and maintenance facilities further constrains convenient use. Nevertheless, the strong motivation to overcome these barriers lies in the ability of this technology to directly translate technical precision into tangible agricultural benefits. The value of research should not be measured solely by improvements in model accuracy but rather by whether these improvements can lead to better field management decisions. Taking leaf area index (LAI) estimation as an example, the value translation pathway is clear: improving LAI estimation accuracy (R²) from 0.75 to 0.90 through more effective feature selection enables the production of a more reliable spatial distribution map of LAI. Using such a high-precision “prescription map,” farm managers can detect subtle within-field variations in crop growth and replace uniform fertilization with variable rate application (VRA). This allows nitrogen inputs to be increased in weaker growth zones while reduced or withheld in vigorous zones, potentially cutting nitrogen use by 15–30 % without sacrificing yield. The same logic applies to variable rate irrigation (VRI), which has the potential to reduce agricultural water use by more than 20 %. Ultimately, this management paradigm—driven by high-precision remote sensing data—can be translated into a virtuous cycle of benefits across three dimensions: economic (reduced input costs), productive (maintained or enhanced yield and quality), and environmental (reduced non-point source pollution).

Therefore, meeting these challenges and unlocking the full value of UAV hyperspectral technology requires a collaborative innovation strategy that integrates technology, business, and policy. On the technological front, the priority is to lower entry barriers by developing more affordable sensors and building cloud-based, automated, one-click data processing platforms. From the perspective of business models, the key lies in promoting a “Technology-as-a-Service” (TaaS) approach, enabling producers to shift from “purchasing assets” to “purchasing services,” thereby converting prohibitive capital expenditures (CAPEX) into more manageable operational expenditures (OPEX). At the policy and ecosystem level, government investment in rural digital infrastructure and subsidies to support technology adoption are essential for creating a favorable application environment and accelerating large-scale deployment. Ultimately, the true value and scalability of any agricultural technology depend on whether it can benefit smallholder farmers, who constitute the overwhelming majority of producers worldwide. Supported by the above strategies, the feasibility of implementation can be evaluated across three dimensions. First, cost feasibility can be achieved through service-based models, freeing smallholders from the need to purchase equipment individually. Second, computational feasibility is ensured by cloud platforms, allowing users to access results with only a smartphone or a basic computer. However, the ultimate determinant of adoption is extreme user-friendliness. The delivery format must be fully “de-professionalized.” For example, packaging outputs into a mobile app that not only uses red–yellow–green maps for early warnings but also provides localized, plain-language recommendations—such as “Red zone: crop may be nitrogen-deficient; immediate fertilization is advised.” Moreover, to build trust, such intelligent diagnostic recommendations should be validated and interpreted through local agricultural extension services. In summary, by building an integrated ecosystem—comprising technology providers, cloud platforms, user-friendly interfaces, and local extension services—it is entirely feasible to transform this cutting-edge technology into an inclusive, practical tool that smallholder farmers can access, afford, and benefit from. This, in turn, would unlock its full potential for strengthening global food security.

8. Conclusion

UAV-based hyperspectral imaging provides unprecedented data dimensionality for the precise monitoring of crop physiological and biochemical parameters. However, its high dimensionality, redundancy, and context sensitivity also present fundamental challenges. One of the central issues is how to effectively extract or select the most representative features from this vast pool of spectral information. This study articulates a core perspective: the effectiveness of any feature extraction or selection method largely depends on the specific agricultural context—defined by a multidimensional decision space encompassing crop type, growth stage, and monitoring objective. Therefore, the selection and evaluation of methods should be grounded within this framework, which enables a more nuanced understanding that their effectiveness fundamentally stems from the degree of alignment between methodological mechanisms and specific agricultural challenges.

First, the intrinsic specificity of different crops represents the primary constraint in determining the applicability of any given method. Both external canopy architecture (e.g., the tall, open structure of maize versus the densely tillering nature of wheat) and internal physiological mechanisms (e.g., leaf anatomical traits, water and nutrient use efficiency) lead to crop-specific spectral responses [170]. As a result, when estimating macro-scale structural parameters such as leaf area index (LAI), simple vegetation indices (VIs)—which are primarily based on general principles of red light absorption and near-infrared reflectance—often exhibit good cross-crop transferability. Examples include the classic Normalized Difference Vegetation Index (NDVI) and the Soil-Adjusted Vegetation Index (SAVI), which is designed for improved soil background correction. However, when the objective shifts to the accurate retrieval of specific internal nutrients (e.g., nitrogen content), the corresponding subtle spectral signals are often masked by extensive redundancy across hyperspectral bands. In such cases, feature selection methods that directly optimize for model prediction accuracy offer a distinct advantage. Second, the crop growth stage introduces a dynamic temporal axis along which method selection must be continuously adapted. An effective strategy involves stage-specific identification and resolution of key spectral challenges throughout the growing season. The general patterns can be summarized as follows: during early growth stages [171], the main challenge lies in interference from background signals due to low canopy coverage. The effectiveness of a method at this stage is largely determined by its ability to suppress background noise. Therefore, vegetation indices or spectral unmixing models that resist soil background interference are often more advantageous at this stage. As the crop enters its vegetative growth phase [172], the primary issue shifts to spectral saturation caused by dense canopy coverage. At this stage, methodological focus should pivot to identifying less saturation-prone spectral regions (e.g., the red-edge) or employing advanced algorithms capable of automatic feature optimization. In the senescence stage [173], the challenge stems from the decoupling of physiological processes, such as nutrient remobilization. At this point, static assumptions about canopy greenness become invalid, and it becomes necessary to construct features that dynamically capture senescence processes, supported by robust nonlinear modeling approaches. This temporally adaptive perspective critically highlights the inherent limitations of applying a single model across the entire growing season. Finally, the complexity of the monitoring objective itself dictates the trade-off between model performance and interpretability. For parameters with clear physical meaning—such as plant water content—simple spectral indices are often preferred due to their transparency and ease of interpretation. For unsupervised exploratory analysis, Principal Component Analysis (PCA) remains a widely used and effective tool. In contrast, when dealing with complex targets such as yield prediction, deep learning models offer superior performance, yet their “black-box” nature and strong dependence on large labeled datasets remain major barriers to practical deployment [170]. In summary, hyperspectral feature engineering is fundamentally a systematic, context-driven decision-making process. The central contribution of this review lies in providing a clear analytical framework and systematic guidance for method selection, tailored to the specific demands of diverse agricultural monitoring scenarios.

This paper provides a systematic review of feature selection and extraction methods in UAV-based hyperspectral technology for crop physiological and biochemical parameter inversion and monitoring. Through a comprehensive literature analysis, we summarized current research progress, with a focus on filter-based, wrapper-based, and embedded feature selection methods, analyzing their applicability, strengths, and weaknesses across different application scenarios. Additionally, the theoretical foundations and practical applications of various feature extraction methods were discussed in detail, providing researchers with a deeper understanding of their value and limitations in crop parameter inversion and monitoring. Furthermore, this paper offers an in-depth review of studies on vegetation indices, discussing the performance of commonly used indices such as NDVI and EVI in estimating nitrogen and chlorophyll content, emphasizing their critical role in extracting crop physiological and biochemical information. A comparative analysis of these indices clarified their importance and applicability in agricultural remote sensing. Finally, this paper has outlined the main challenges of existing methods and provided perspectives for future research. Optimizing feature engineering is not merely a technical refinement, but represents a paradigm shift toward greater intelligence, automation, and multimodal integration. The integration of physical models with deep learning, together with the development of advanced algorithms capable of automatically learning domain-invariant features and deeply fusing spectral–spatial information, will be crucial for enabling UAV-based hyperspectral technology to achieve large-scale, inclusive applications in precision agriculture.

Author contributions

Liuchang Xu was responsible for conceptualization, project administration, and resource collection. Luyao Chen contributed to data curation, formal analysis, investigation, methodology, visualization, and writing the original draft. Hailin Feng and Liuchang Xu provided funding support and supervision. Hailin Feng also contributed to validation and reviewing & editing the manuscript. Jianqin Huang, Ketao Wang, Zijia Yang, Xiang Weng, Kai Fang, and Qianqian Luo participated in supervision, validation, and reviewing & editing. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the “Pioneer” and “Leading Goose” R&D Program of Zhejiang (Grant Nos. 2022C02009 and 2022C02044); and the National Natural Science Foundation of China (Grant Nos. 62102366 and 62403433).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article: