Growth-stage deep learning for UAV hyperspectral estimation of leaf nitrogen content in winter wheat

Abstract

Leaf nitrogen content (LNC) is critical for crop nutrition and precision fertilization. UAV-based hyperspectral sensing enables rapid field-scale nitrogen diagnosis in winter wheat, but stage-dependent drift in the LNC-spectral relationship limits multi-temporal monitoring. To address this limitation, this study proposed a Growth-stage Deep learning model for Leaf Nitrogen Content estimation in winter wheat (GD-LNC). The core innovation of GD-LNC lies in the synergistic integration of explicit growth-stage embedding with a dynamic feature fusion gating mechanism. The growth stage was incorporated as a key parameter in the deep learning-based hyperspectral model for LNC estimation. This strategy improved the generalization ability of multi-temporal LNC estimation during the grain-filling period. This study utilized the datasets from two growing seasons of multi-factor winter wheat experiments. The datasets included UAV-based hyperspectral images and synchronous ground samples. LNC-sensitive bands were selected via stepwise projection algorithm (SPA) and shuffled frog leaping algorithm (SFLA).

Subsequently, a multi-temporal dynamic feature extraction and growth-stage embedding branch was constructed. An adaptive fusion mechanism was then employed to dynamically adjust the contribution weights of each feature. These components together enabled robust estimation and mapping of LNC. Results show hyperspectral features strongly responded to LNC, and SFLA-selected bands retained key spectral information, improving estimation accuracy and stability. GD-LNC outperformed PLSR, LSTM and RF, with SFLA-GD-LNC achieving R² of 0.88 and RMSE of 0.20% (training), and R² of 0.85 with RMSE 0.22% (validation), and R² of 0.88 with RMSE 0.21% (test). The model captured spatial and temporal LNC dynamics, providing a generalizable UAV hyperspectral framework for nitrogen monitoring and cross-stage adaptation.

1. Introduction

Nitrogen is an essential nutrient for crop growth and development. It plays a crucial role in photosynthesis, substance synthesis, and metabolic regulation. Nitrogen also directly affects yield formation and quality improvement [1,2]. As a typical high-nitrogen-demanding crop, wheat's growth and development are highly sensitive to plant nitrogen status [3]. Leaf nitrogen content (LNC) is a key indicator of crop nitrogen status. Compared with soil nitrogen measurements, LNC assessment is more convenient and less costly. This makes LNC suitable for rapid field-level diagnosis [4]. However, the ability to quickly and accurately obtain LNC at the field scale remains a critical limitation for precision fertilization management. Therefore, developing rapid and reliable nitrogen diagnostic methods is of great significance for enabling precise topdressing and improving nitrogen use efficiency [5].

The advancement of remote sensing technologies has provided a new approach for low-cost, non-destructive monitoring of crop nitrogen status. Proximal sensing offers advantages in mitigating the effects of sunlight changes and soil background interference. However, its spatial coverage is limited and it is difficult to meet monitoring requirements at the regional scale. Unmanned aerial vehicle (UAV)-based remote sensing is characterized by high maneuverability and high spatiotemporal resolution. It has been widely applied for the inversion of parameters such as leaf area index [6], chlorophyll content [7], biomass [8], and yield [9]. In terms of sensor level, multispectral imagery offers high processing efficiency. It has fewer bands, making it difficult to effectively characterize the subtle nitrogen-related spectral responses. In contrast, UAV-based hyperspectral imagery has the advantage of continuous narrow bands [10]. It enables more refined detection of nitrogen-sensitive spectral features. This approach shows greater potential for improving inversion accuracy and elucidating spectral-agronomic trait relationships [11,12].

In terms of crop nitrogen inversion methods, machine learning models have demonstrated strong capability in exploiting high-dimensional spectral information for nonlinear fitting [13]. Previous studies have achieved robust prediction of plant nitrogen concentration through spectral feature optimization combined with random forest modeling. These approaches show superior performance to linear models in reducing estimation errors [14,15]. In recent years, deep learning has shown strong end-to-end feature learning ability. It can more fully depict the complex nonlinear relationship between remote sensing data and crop traits. Deep learning has been gradually applied to the task of crop nitrogen state identification and estimation [16,17]. Furthermore, recent studies have validated the effectiveness of combining spectral information with machine learning for plant biochemical trait prediction and substance classification [18,19]. Deep learning has also demonstrated strong potential in automated plant phenotyping tasks [20]. Meanwhile, the long short-term memory (LSTM) model, a specific variation of recurrent neural networks with short-term memory capability. Compared with traditional RNNs, LSTM introduces gated memory mechanisms that enable controlled information flow. It is therefore more suitable for processing and predicting events with long temporal dependencies or delayed effects in time series. The network solves problems such as gradient disappearance and gradient explosion caused by time backpropagation during training. Wheat production can be accurately forecast using an LSTM model [21]. Previous studies have demonstrated that LSTM performs effectively in various crop management tasks, including nitrogen nutrition diagnosis [22]. Convolutional neural networks can extract discriminative features through hierarchical structures [23]. The Transformer structure utilizes the self-attention mechanism to enhance global feature expression ability [24]. Related studies have further confirmed the feasibility and effectiveness of deep learning for leaf nitrogen content prediction [25].

During the grain-filling stage, wheat undergoes substantial nitrogen remobilization from vegetative organs—particularly leaves—to developing grains. Approximately 60-80% of grain nitrogen is derived from pre-anthesis reserves and redistributed through leaf senescence [26]. Consequently, leaf nitrogen content (LNC) during this period serves as a direct indicator of the plant's nitrogen source capacity and the progression of nitrogen reallocation. Moreover, leaf nitrogen is tightly coupled to photosynthetic capacity. Rubisco alone accounts for approximately 25-30% of total leaf nitrogen [27]. Maintaining adequate LNC during grain filling is therefore essential for sustaining photosynthetic activity and delaying premature leaf senescence. It also ensures sufficient carbon assimilation to support grain development. A rapid decline in LNC leads to early senescence and a shortened photosynthetically active duration. This ultimately reduces grain weight and yield [28]. In addition, LNC during grain filling is directly and strongly associated with grain protein concentration, a key quality parameter for wheat end-use [29]. This stage also represents the final window for in-season nitrogen management decisions, such as late-season foliar nitrogen application. These practices aim to improve grain protein content and nitrogen use efficiency [5]. Furthermore, LNC during grain filling is highly sensitive to spatial variability in soil nitrogen supply, water availability, and microclimatic conditions. Its spectral signal is also affected by changes in canopy structure, leaf moisture, and chlorophyll degradation [30]. Traditional destructive sampling methods are labor-intensive and impractical for capturing such spatial heterogeneity and temporal dynamics at field scale. However, most existing remote sensing-based LNC estimation studies have focused on early to mid-season stages. The grain-filling stage remains comparatively under-represented, despite its importance for final yield and quality formation. This represents a significant knowledge gap. Therefore, developing accurate and stage-adaptive LNC estimation approaches for the grain-filling period is of great significance for UAV-based hyperspectral nitrogen monitoring and precision fertilization management.

There are still challenges in improving the UAV-based estimation accuracy by utilizing the temporal variation patterns of LNC during the growth process of wheat. During the grain-filling period, nutrients are rapidly transferred from leaves to grains. The LNC at different stages shows significant dynamic changes. It is easily influenced by environmental conditions, light, temperature, and moisture [30]. Consequently, models trained on a single observation date are difficult to generalize to other key stages within the grain-filling period, leading to prediction accuracy that varies with the observation date [31]. Meanwhile, hyperspectral data have the characteristics of high dimensionality, strong redundancy, and complex correlation. The traditional convolutional structure focuses on local modeling and has difficulty fully expressing the global dependency relationships between bands. The local receptive fields and translation equivariance assumptions inherent in CNNs are not sufficient for capturing holistic image-level semantics [32]. Pure Transformer architectures, which use global self-attention, can achieve superior performance on vision tasks. Likewise, conventional convolutional architectures are constrained by their fixed receptive fields when modeling long-range dependencies in sequential data. State space models such as Mamba provide an efficient alternative by enabling scalable sequence modeling and effectively capturing long-range temporal dependencies [33]. Some cascaded convolutional networks may introduce repetitive feature extraction and computational redundancy. Moreover, deep residual networks have been shown to behave as ensembles of exponentially many shallow paths. Removing individual layers has little impact on performance, suggesting that much of the depth-wise computation in cascaded architectures is redundant [34]. Furthermore, it has been demonstrated that large dense networks contain sparse subnetworks — termed “winning tickets.” These pruned subnetworks can retain accuracy comparable to that of the original full network, indicating that a substantial proportion of the filters, parameters, and computations in cascaded convolutional architectures may be redundant [35]. This further limits their efficiency and generalization capability in multi-temporal continuous inversion tasks [36]. Therefore, it is necessary to construct a modeling framework with stage adaptability for the dynamic changes during the grain-filling period to enhance the cross-phase stability of LNC estimation.

To address the above issues, this study proposes a Growth-stage Deep learning model for Leaf Nitrogen Content estimation in winter wheat (GD-LNC). Within a 1D CNN-based spectral encoder inspired by the U-Net encoder structure, the model incorporates a Transformer self-attention mechanism to model global dependencies among hyperspectral bands. This model integrates multi-stage temporal information acquired during the grain-filling period. Specifically, it employs a dynamic weighting mechanism to assign differential importance to nitrogen-sensitive features across different temporal stages. The purpose of this design is to account for the temporal variability in spectral responses associated with nitrogen accumulation dynamics during grain filling. This helps construct a more robust feature representation for multi-temporal leaf nitrogen content inversion. The specific objectives of this study are to: (1) apply the SPA and SFLA algorithms to select nitrogen-sensitive hyperspectral bands and verify their effectiveness; (2) integrate hyperspectral data with growth-stage information to construct a dynamic deep feature fusion framework, enabling adaptive regulation of multiple feature contributions; and (3) develop and evaluate the GD-LNC model to achieve continuous and stable estimation of LNC during the grain-filling period of winter wheat. Compared with existing multi-temporal hyperspectral approaches that address temporal variability through either external stage labeling or static model architectures, the novelty of this study lies in the explicit coupling of growth-stage embedding with dynamic feature fusion gating. Neither mechanism alone is sufficient on its own. The growth-stage embedding provides phenological context for spectral interpretation, while the fusion gating mechanism ensures adaptive recalibration of feature contributions. Their combination enables robust cross-stage generalization that neither component could achieve independently.

2. Materials and methods

2.1. Experimental design

From 2022 to 2024, a two-season variable water-fertilizer control experiment was conducted consecutively at the Xinji Experimental Station of Hebei Agricultural University in China (37°59′24″ N, 115°12′5″ E). A multifactorial experimental design was employed, comprising four irrigation levels and five nitrogen application levels. Irrigation was determined using the Cropwat model [37] to calculate crop water requirements (CWR). The experiment includedfour irrigation treatment groups: normal irrigation (W100, providing 100% of CWR), mild water stress (W75, 75% of CWR), moderate waterstress (W50, 50% of CWR), and a control group (W0), relying solely onnatural rainfall. Irrigation was initiated when CWR reached 30 mm. The nitrogen treatments ranged from N0 to N4, corresponding to application rates of 0, 180, 240, 300, and 360 kg ha⁻¹. The experimental field covered an area of 0.3 ha, with 1 m-wide buffer rows surrounding the field and 0.5 m-wide observation alleys set between plots. During the 2022-2024 experimental period, five winter wheat cultivars were used, including Gaoyou 2018, Malan 1, Hengguan 35, Shiluan 02-1, and Jimai 22. Different water-nitrogen and cultivar interaction experiments were conducted during the 2022-2023 and 2023-2024 growing seasons. In 2022-2023, four cultivars were employed: Gaoyou 2018, Malan 1, Hengguan 35, and Shiluan 02-1. In 2023-2024, two cultivars were used: Gaoyou 2018 and Jimai 22.

During the 2022-2023 growing season, three types of experiments were conducted. First, a water-nitrogen interaction experiment was performed using the cultivar Gaoyou 2018. The experiment included 20 water-nitrogen treatments with three replicates, resulting in 60 plots. Second, a cultivar-water interaction experiment was conducted using the cultivars Malan 1, Hengguan 35, and Shiluan 02-1. Third, a cultivar-nitrogen interaction experiment was also conducted using the same three cultivars. These two experiments included a total of 27 treatments with three replicates, resulting in 81 plots. Four observation campaigns were carried out during the grain-filling stage, yielding a cumulative total of 564 samples. During the 2023-2024 growing season, the cultivars Gaoyou 2018 and Jimai 22 were used in a cultivar-water-nitrogen interaction experiment. Two cultivars (Gaoyou 2018 and Jimai 22) were combined with 4 irrigation levels and 5 nitrogen levels, yielding 40 treatments. With three replicates, the experiment comprised 120 plots in total. Four observations were conducted during the grain-filling stage, yielding a total of 480 samples, as shown in Table 1.

Table 1.

Summary of the experimental design, treatment structure, plot number, observation dates, and sample size during the 2022–2024 growing seasons.

Growing season	Experimental module	Cultivar(s) and treatment structure	No. of plots	Observation dates	Sample size
2022–2023	Water–nitrogen interaction experiment	Gaoyou 2018; 4 irrigation levels × 5 nitrogen levels	60	4	240
2022–2023	Cultivar–water interaction experiment	Malan 1, Hengguan 35, and Shiluan 02-1; 3 cultivars × 4 irrigation levels	36	4	144
2022–2023	Cultivar–nitrogen interaction experiment	Malan 1, Hengguan 35, and Shiluan 02-1; 3 cultivars × 5 nitrogen levels	45	4	180
2022–2023	Subtotal	Four cultivars were used in three experimental modules	141	4	564
2023–2024	Cultivar–water–nitrogen interaction experiment	Gaoyou 2018 and Jimai 22; 2 cultivars × 4 irrigation levels × 5 nitrogen levels	120	4	480
2022–2024	Total	Five cultivars were used across two growing seasons	261	8 observation campaigns	1044

Experimental procedures. All experimental plots were arranged in a randomized complete block design with three replicates. Each plot measured 5 m (east-west) × 4 m (north-south), with an area of 20 m². Winter wheat was sown at a basic seeding rate of 3.75 million plants ha⁻¹. Irrigation was applied through a surface drip irrigation system. Water meters were installed in each plot to precisely control irrigation volume. The irrigation amount for each treatment was calculated based on crop water requirements (CWR). The W100 treatment received 100% of CWR (full irrigation). W75 received 75% of CWR (mild drought stress). W50 received 50% of CWR (moderate drought stress). W0 received no irrigation and relied solely on natural precipitation. Water stress was imposed during two growth stages: from regreening to jointing and from flowering to 25 days after flowering. Full irrigation was maintained from jointing to flowering in all irrigated treatments. Nitrogen fertilizer was applied in two split applications: 50% as a basal dressing at sowing and 50% as a topdressing during the growing season. Other field management practices were carried out consistently with local conventional high-yield field practices, as illustrated in Fig. 1, Fig. 2.

Fig. 1.

Study area and distribution of experimental plots.

Fig. 2.

Flow chart of experimental.

2.2. Data acquisition

The growth and developmental status of winter wheat was assessed using the Zadoks scale [38], with key phenological stages recorded when 50% of the plants within a population reached a given stage. The specific measuring dates are listed in Table 2. On the corresponding dates, UAV-based hyperspectral canopy imagery was acquired concurrently with measurements of leaf nitrogen content. Subsequently, plot boundary vector data were spatially overlaid with the hyperspectral imagery to extract the mean hyperspectral reflectance for each plot.

Table 2.

Data acquisition schedule.

Data	2022-2023	Zadoks scale	2023-2024	Zadoks scale
UAV Hyperspectral images, LNC	8 May	Z73	2 May	Z71
	13 May	Z75	8 May	Z73
	25 May	Z83	14 May	Z75
	1 Jun	Z91	27 May	Z85

Z71: Caryopsis water ripe; Z73: Early milk; Z75: Medium milk; Z83: Early dough; Z85: Soft dough; Z91: Caryopsis hard.

2.2.1. Ground data acquisition

Ground sampling was conducted following a random principle, whereby representative and pest- and disease-free wheat flag leaves were selected from each treatment plot. Sampling time, growth stage, and sampling location were recorded simultaneously, and the samples were sealed, labeled, and stored. Leaf samples were first heat-killed at 105 °C for 30 min and then oven-dried to constant weight at 65-80 °C. The dried samples were ground and passed through a 100-mesh sieve. A 0.1 g sample of the dried material was digested using a sulfuric acid-hydrogen peroxide method. After cooling, the digest was diluted to a final volume of 50 mL. Following clarification, 1 mL of the supernatant was further diluted to 5 mL for subsequent determination.

Leaf nitrogen concentration (LNC) was determined by measuring the NH₄⁺ content in the digested solution using a continuous flow analyzer. Absorbance was quantified at a specific wavelength based on a colorimetric reaction, and the nitrogen concentration (C) was calculated from a standard calibration curve. Because the instrument output represents nitrogen concentration in solution, LNC was calculated by accounting for the final digestion volume, dilution factor, and sample mass; the parameters and calculation formula are provided in Equation (1).

$$\text{LNC}=\frac{C\times V\times D}{M}$$ (1)

In the equation, LNC denotes leaf nitrogen concentration (g·kg⁻¹), C represents the nitrogen concentration measured by the continuous flow analyzer (mg·L⁻¹), V is the final digestion volume (L; 0.05 L in this study), D is the subsequent dilution factor (5 in this study), and m is the mass of the weighed dry sample (kg; 0.0001 kg in this study).

2.2.2. Acquisition of UAV-based hyperspectral imagery

The images of the experimental area were acquired using a DJI M300 RTK unmanned aerial vehicle equipped with an X20P hyperspectral imaging sensor. The X20P sensor is fitted with a high-performance 20-megapixel CMOS detector, and its detailed parameters are provided in Table 3.

Table 3.

Specific parameters of UAV-based hyperspectral imaging system.

UAV		Hyperspectral camera
Model	DJI M300 RTK	Model	X20P
Flight time	11:00-13:00	Spectral range	350-1000 nm
Flight altitude	30m	Band width	4 nm
Flight speed	2 m/s	Wavelength accuracy	0.5 nm
Forward overlap	80%	Number of bands	164
Lateral overlap	80%	Pixel	20 million

2.2.3. Spectral reflectance acquisition and preprocessing

The raw hyperspectral imagery was processed using Agisoft Metashape Professional software for stitching, geometric correction, and radiometric calibration to generate high-resolution images containing spectral reflectance information. Systematic spectral calibration was performed on the hyperspectral data to ensure radiometric consistency and cross-temporal comparability. Standard white panels (Spectralon) were used as references. White reference reflectance and dark current signals were collected before and after each observation. Reflectance conversion was then conducted using the following formula (2):

$$R=\frac{{DN}_{target}-{DN}_{dark}}{{DN}_{white}-{DN}_{dark}}$$ (2)

Among them, $D{N}_{target}$ represents the original digital signal of the target object, $D{N}_{white}$ is the reference signal of the whiteboard, and $D{N}_{dark}$ is the dark current signal. The calibrated hyperspectral data were then smoothed using the Savitzky-Golay filter to reduce noise while preserving the spectral curve shape [39]. Using the plot boundary data, the mean spectral reflectance for each band within each plot was calculated for subsequent feature extraction and modeling analyses.

2.3. Methods

2.3.1. Selection of sensitive spectral bands

Hyperspectral full-band data contain a certain degree of redundant information, which can interfere with feature representation and reduce the accuracy of LNC estimation. To improve data processing efficiency and enhance the capability of features to characterize the dynamic changes of winter wheat LNC during the grain-filling stage, sensitive spectral bands were selected from the hyperspectral data using successive projections algorithm (SPA) and shuffled frog leaping algorithm (SFLA). SPA minimizes band collinearity through forward iteration, selecting candidate bands with low redundancy from the original spectrum. For SFLA, the absolute Pearson correlation coefficient between each spectral band and LNC was first calculated. It was used as an auxiliary sensitivity indicator. This indicator was used to evaluate the individual response of each wavelength to LNC. However, Pearson correlation was not used as the formal optimization objective. During the SFLA optimization process, each frog represented a candidate wavelength subset. The fitness of each subset was evaluated using the cross-validated RMSE of the LNC estimation model. This model was constructed with the selected wavelengths. Therefore, the objective of SFLA was to identify the wavelength subset with the lowest cross-validated RMSE.

SPA is a forward variable selection method that effectively reduces collinearity among variables [40] and can be applied to select characteristic wavelengths for quantitative spectral analysis [41]. SFLA is a population-based intelligent optimization method. It combines local search with global information exchange. This enables rapid identification of wavelength combinations that are more sensitive to the target variable and exhibit lower redundancy across the full spectral range. It thereby improves both the accuracy and efficiency of spectral analysis [42].

To ensure reproducibility and methodological transparency, the key hyperparameter settings for both algorithms are specified as follows. For SPA, the maximum number of selected wavelengths was set to 10, and the optimal number of wavelengths was determined by minimizing the root mean square error (RMSE) on the calibration set through a systematic forward selection procedure. For SFLA, the population size was set to 300, the number of memeplexes was set to 15, the maximum number of global iterations was set to 200, and the maximum number of local search iterations within each memeplex was 30. The number of selected bands was constrained to the range of 10-20, and the final band combination was determined by minimizing the cross-validated RMSE of the LNC estimation model. These parameter configurations were determined through preliminary experiments conducted only on the training set and were guided by recommended ranges in the literature [40,42]. The independent validation set was not used during this process.

To avoid information leakage, all parameter determination and sensitive-band selection procedures for SPA and SFLA were performed only on the training set. The independent validation set was not involved in SPA forward selection, SFLA fitness evaluation, parameter tuning, or the determination of the final wavelength subset. For SPA, the optimal number of wavelengths was determined by minimizing RMSE within the training set. For SFLA, each candidate wavelength subset was evaluated using cross-validated RMSE calculated only from the training samples. After the optimal wavelength subsets were identified from the training set, the selected bands were fixed. They were then applied unchanged to the independent validation set for final model evaluation.

2.3.2. Construction of the growth-stage deep learning model (GD-LNC)

To address the intrinsic complexity of multi-source data in terms of temporal dynamics and information complementarity. A Growth-stage Deep learning model for Leaf Nitrogen Content estimation in winter wheat (GD-LNC) was developed. The model is built upon a hybrid architecture that combines a Transformer with a 1D CNN-based spectral encoder, and consists of three main components: a source feature extraction module, a dynamic feature deep fusion module, and a regression prediction module. First, discriminative initial features are extracted from different input sources. Subsequently, a dynamic deep fusion module based on multi-head self-attention and adaptive weight allocation is introduced. It adaptively integrates feature representations from different sources or hierarchical levels. This enhances the integration and selection of key information. Finally, a regression prediction module performs end-to-end estimation of LNC based on the fused deep features, as illustrated in Fig. 3.

Fig. 3.

Overall architecture of the GD-LNC model.

2.3.2.1. Source feature extraction module

This module adopts a parallel dual-stream architecture to separately process hyperspectral imagery data and growth-stage data, as shown in Fig. 4. Specifically, in the hyperspectral branch, an improved 1D CNN-based spectral encoder inspired by the U-Net encoder structure is employed, in which feature extraction is performed using a dilated temporal convolutional network (Dilated temporal CNN) composed of three one-dimensional convolutional layers (Conv 1d-1, Conv 1d-2, and Conv 1d-3). After multiple one-dimensional convolutional layers and nonlinear activations, temporal spectral features are derived through feature integration and compression. In the growth-stage branch, an improved Transformer-based multi-head self-attention mechanism is employed, with multi-temporal information during the grain-filling period serving as the input. In this study, growth stages were encoded on a per-observation-time basis. In this study, growth-stage information was defined according to the Zadoks scale. UAV hyperspectral observations and corresponding LNC measurements were conducted at the actual grain-filling stages recorded in Table 1, including Z73, Z75, Z83, and Z91 in the 2022–2023 growing season, and Z71, Z73, Z75, and Z85 in the 2023–2024 growing season. Therefore, artificial temporal labels such as T1–T8 were not used. Instead, each sample was assigned a categorical growth-stage label based on its actual Zadoks code at the time of data acquisition. The same Zadoks stage observed in different growing seasons was encoded as the same biological growth-stage category. These categorical Zadoks labels were converted into integer indices and then transformed into learnable embedding vectors through an embedding layer. Calendar dates, growing-season labels, and observation campaign indices were not used as growth-stage inputs in the GD-LNC model. The resulting growth-stage embeddings were concatenated with hyperspectral and structural features before feature fusion. Both feature streams were processed through ReLU activation functions and nonlinear transformations (Conv1d). The extracted features were then concatenated and passed through a linear projection layer to produce a unified feature representation, which serves as the basis for subsequent feature fusion.

Fig. 4.

Architecture of the feature extraction module in the GD-LNC model.

As a result, the model can configure appropriate feature extractors for different growth stages. The expansion factor increases exponentially with the progression of the growth period. This aligns with the gradual temporal changes in wheat canopy spectra. It enables the network to adaptively adjust the distribution shifts of spectral features to accommodate variations across different growth stages.

2.3.2.2. Dynamic deep feature fusion module

This module serves as the core of the network, where preliminarily processed features are adaptively fused through a dynamic fusion mechanism, as illustrated in Fig. 5. The mechanism consists of two complementary components: a global feature interaction layer and an adaptive weight allocation layer.

Fig. 5.

Overall architecture of the GD-LNC dynamic deep feature fusion and regression prediction modules.

In the global feature interaction layer, the concatenated joint feature $f_{combined}\in R_{88}$ (formed by 64-dimensional hyperspectral features, 16-dimensional structural features, and 8-dimensional growth-stage embedding vectors) is fed into a two-layer Transformer decoder. Each decoder layer comprises eight multi-head attention heads (nhead = 8) with a hidden dimension of 88. The joint feature serves simultaneously as the Query, Key, and Value input. Through multi-head self-attention, the decoder explicitly models global dependencies and interactions among the multi-source features, producing an output feature $f_{output}\in R_{88}$ that encodes richer contextual information.

In the adaptive weight allocation layer, an MLP-based weight generation network takes the 88-dimensional joint descriptor as input (64 + 16 + 8 = 88) and passes it through a two-layer fully connected network (88 → 32→3). A Softmax function then generates three dynamic fusion weights $w_1$, $w_2$, $w_3$, corresponding to the hyperspectral, structural, and growth-stage embedding features, respectively. The final deeply fused feature representation is obtained through weighted summation: $f_{fused}=w_1.h_{spectral}+w_2.h_{struct}+w_3.h_{period}$

Through this dual mechanism, the model can explicitly capture global inter-feature dependencies while adaptively adjusting the relative importance of each feature source across different growth stages. Thereby enhancing the integration and selection of key information for LNC estimation.

2.3.2.3. Regression prediction module

This module receives the dynamically fused high-level features and performs the final prediction task, as illustrated in Fig. 5. The fused features first pass through a fully connected layer for further nonlinear transformation and feature integration, followed by a ReLU activation function to enhance the model's expressive capability. The transformed features are mapped to the prediction target dimension through a linear layer. They are then passed through a Sigmoid activation function to produce the final regression prediction (Output). This completes the end-to-end process from multi-source data input to the generation of specific prediction results.

2.4. Validation

To comprehensively evaluate the predictive performance of the proposed model for winter wheat LNC and compare it with traditional and deep learning methods, random forest regression (RF), partial least squares regression (PLSR), and long short-term memory (LSTM) networks were used as reference models to benchmark the growth-stage deep learning model GD-LNC. PLSR alleviates the multicollinearity effect of hyperspectral features by decomposing and reconstructing the set of independent variables [43]. RF is a nonparametric ensemble method based on decision trees. It reduces the impact of variable correlation on model stability through random selection of feature subsets, thereby exhibiting strong nonlinear fitting capability [44,45]. LSTM, proposed by Hochreiter et al. [46], is an extension of recurrent neural networks (RNNs), is designed to address the gradient vanishing and explosion problems encountered in traditional RNNs. Unlike conventional RNNs, LSTM employs gated memory units, including forget gates, input gates, and output gates. t These gates selectively retain or discard information during training, thereby improving the modeling capability for sequential data.

The hyperparameters of the reference models were specified as follows. For RF, the number of trees (n_estimators) was set to 50 to balance model complexity and computational cost. The maximum tree depth (max_depth) was limited to 6 to prevent overfitting. The minimum number of samples required to split an internal node (min_samples_split) was set to 10, and the minimum number of samples at a leaf node (min_samples_leaf) was set to 5. These settings were used to enhance model generalization and smooth prediction outputs. The proportion of features considered at each split (max_features) was set to 0.5 to introduce randomness and improve model robustness. The random seed (random_state) was fixed at 42 to ensure reproducibility. This parameter combination was designed to control model complexity while maintaining prediction accuracy and generalization performance. For PLSR, the dataset was split into 70% training, 15% validation, and 15% test subsets with a fixed random seed (random_state = 42). The input variables consisted of the selected hyperspectral features, and the output variable was LNC. The number of latent components (n_components) was determined through five-fold cross-validation on the training set. The search range for n_components was set from 1 to 10. The optimal number of components was selected based on the maximum cross-validated R², computed using the cross_val_predict function in scikit-learn. RMSE was also calculated at each component level to comprehensively evaluate model performance. The final PLSR model was re-fitted on the full training set using the optimal number of components and subsequently evaluated on both the training and independent validation sets. No standardization or normalization was applied to the input data, and all other PLSR parameters were left at their scikit-learn defaults. All model implementations were carried out using the scikit-learn library (version 1.3.2) in Python 3.10.

An LSTM network was developed as a temporal recurrent baseline for LNC estimation. For each experimental plot, UAV hyperspectral observations were collected at multiple dates during the grain-filling period. These observations were arranged in chronological order. Therefore, the sequence dimension of the LSTM input represented successive observation dates or growth stages. It did not represent different wavelengths. At each time step, the input consisted of the selected hyperspectral features corresponding to that observation date. In this way, the LSTM model was used to learn the temporal evolution of hyperspectral responses associated with LNC changes across the grain-filling period.

The input features and target values were normalized to [0, 1] using MinMaxScaler. The LSTM model consisted of two LSTM layers. Each layer had 64 hidden units. The two LSTM layers were followed by a fully connected output layer for LNC prediction. The model was trained using the Adam optimizer. The learning rate was 0.001. The batch size was 32. Mean squared error was used as the loss function. R² and RMSE were used as evaluation metrics.

For GD-LNC, to ensure objective and stable evaluation, this study combined multi-temporal LNC samples from the grain-filling stages across two consecutive years for modeling, yielding a total of 1044 samples (n = 1044). Specifically, PLSR, RF, LSTM, and the proposed GD-LNC model were all trained, validated, and tested using the same 70% training, 15% validation, and 15% test split. Stratified sampling based on growth-stage labels was applied to maintain balanced stage distributions across the subsets. The training set was used for model parameter learning, while the validation set was employed to evaluate model performance and select the optimal model. The independent test set was used only for final performance evaluation.

During training, the Adam optimizer was adopted with a learning rate of 1 × 10⁻⁴ and a batch size of 16. The model was trained for 60 epochs using the mean squared error (MSE) loss function. The model with the highest validation R² was retained for final testing. Model performance was quantitatively evaluated using the coefficient of determination (R²) and the root mean square error (RMSE), as defined in Equations (3) and (4), respectively.

$$R^2=1-\frac{{{\sum }_{i=1}^n\left(Y_i-Y_i\right)}^2}{{{\sum }_{i=1}^n\left(Y_i-Y_i\right)}^2}$$

$$RMSE=\frac{1}{n}{{\sum }_{i=1}^n\left(Y_i-Y_i\right)}^2$$

In these equations, (yi) denotes the measured value, (ŷi) the estimated value, (ȳi) the mean of the measured values, and (n) the number of samples. R² reflects the model's ability to explain the variance of observations, and RMSE reflects the absolute level of prediction error. Together, these metrics are used to compare the fitting and generalization performance of different models on the training and validation datasets.

3. Results and analysis

3.1. Analysis of LNC dynamic changes

Fig. 6 illustrates the distribution characteristics of winter wheat leaf nitrogen content across different experimental treatments in 2023 and 2024 as it varies with measuring dates, where IQR denotes the interquartile range. Measuring dates are encoded using the Zadoks scale, covering Z71, Z73, Z75, Z83, Z85, and Z91, which mainly correspond to the grain-filling stage and its adjacent key phenological phases.

Fig. 6.

Descriptive plots of wheat LNC at different growth stages. Note: LNC, leaf nitrogen content; IQR, interquartile range; Z71: caryopsis water ripe; Z73: early milk; Z75: medium milk; Z83: early dough; Z85: soft dough; Z91: caryopsis hard. (a) Descriptive plot of wheat LNC at growth stages in 2023; (b) Descriptive plot of wheat LNC at growth stages in 2024.

Overall, LNC exhibited a continuous declining trend over the grain-filling period in both seasons. At the Z71 and Z73 stages, LNC distributions were relatively concentrated and stable, whereas a pronounced decrease was observed after entering Z75, with overall levels further declining at Z75 and subsequent stages.

It was observed that there were significant differences in the absolute levels and dispersion of LNC were observed between the two seasons. In 2024, measurements began earlier, starting at Z71, with overall lower LNC levels and smaller variability. LNC values at each stage were mainly concentrated within the range of 1-2. In contrast, observations in 2023 started at a relatively later stage, but overall LNC levels were higher and inter-stage differences were more pronounced. LNC exhibited greater dispersion during the early stages in 2023, gradually declined with the progression of grain filling, and became more concentrated at later stages.

3.2. Selection of LNC-sensitive spectral bands

To reduce redundancy and mitigate the effects of multicollinearity in full-band hyperspectral data, this study applied SPA and SFLA to select sensitive spectral bands from winter wheat leaf hyperspectral data. The results indicate that the feature bands selected by both methods were highly significantly correlated with LNC, with p-values less than 0.01.

SPA selected 10 feature wavelengths at 402, 410, 450, 518, 738, 898, 926, 934, 958, and 982 nm, covering the visible violet-blue, green, red edge, and near-infrared ranges, with an overall discrete distribution. In contrast, SFLA selected 15 feature wavelengths at 494, 498, 502, 506, 510, 662, 666, 678, 682, 686, 770, 778, 810, 994, and 1002 nm, showing a more concentrated distribution across the visible blue-green, red edge, and near-infrared ranges. Compared with SPA, SFLA selects wavelength combinations with lower information redundancy. This helps mitigate multicollinearity among hyperspectral variables. It also provides more discriminative feature inputs for subsequent model development. The wavelength selection results from both methods are summarized in Table 4.

Table 4.

Leaf spectral characteristic bands screened by different algorithms.

Selection methods	Selected wavelength (nm)
SPA	402 410 450 518 738 898 926 934 958 982
SFLA	494 498 502 506 510 662 666 678 682 686 770 778 810 994 1002

3.3. Baseline and ablation analysis of the proposed GD-LNC model

To verify the contribution of growth-stage information and the dynamic feature fusion mechanism, baseline and ablation experiments were conducted. All models were trained and evaluated using the same data partition, input spectral features, preprocessing procedure, and evaluation metrics. The tested configurations included pooled baseline models without growth-stage information, stage-specific baseline models, GD-LNC without growth-stage embedding, GD-LNC without the dynamic fusion gate.

3.3.1. Pooled baseline without growth-stage information

The all baseline models were trained using samples from all growth stages. No growth-stage information was included, and only spectral features were used for LNC estimation. As shown in Table 5, RF achieved the best test performance, with an R² of 0.71 and an RMSE of 0.30%. LSTM showed a similar accuracy, with an R² of 0.70 and an RMSE of 0.32%. PLSR performed worse, with an R² of 0.54 and an RMSE of 0.40%.

Table 5.

Pooled baseline without growth-stage information.

Modeling method	Training set		Validation set		Test set
	R2	RMSE (%)	R2	RMSE (%)	R2	RMSE (%)
LSTM	0.82	0.25	0.74	0.29	0.70	0.32
PLSR	0.61	0.37	0.63	0.34	0.54	0.40
RF	0.84	0.24	0.78	0.26	0.71	0.31

Compared with the models using growth-stage information in Table 8, all pooled baseline models showed lower accuracy. This indicated that growth-stage variation affected the LNC–spectral relationship. Therefore, spectral features alone were insufficient to accurately estimate LNC across the whole growth season.

Table 8.

Comparison of modeling results for two sensitive band selection methods and three modeling algorithms.

Variable screening	Modeling method	Training set		Validation set		Test set
		R2	RMSE (%)	R2	RMSE (%)	R2	RMSE (%)
SPA	PLSR	0.58	0.39	0.56	0.36	0.49	0.41
	LSTM	0.81	0.26	0.65	0.32	0.60	0.36
	RF	0.74	0.31	0.67	0.31	0.62	0.35
	GD-LNC	0.81	0.26	0.79	0.27	0.72	0.30
SFLA	PLSR	0.83	0.25	0.77	0.27	0.79	0.27
	LSTM	0.87	0.21	0.82	0.24	0.83	0.24
	RF	0.86	0.22	0.80	0.26	0.79	0.27
	GD-LNC	0.88	0.20	0.85	0.22	0.88	0.21

3.3.2. Stage-specific baseline models

Stage-specific baseline models were constructed for each growth stage. A separate model was trained for each stage. The prediction results from all stages were then aggregated. The overall performance across the whole growth cycle was calculated. As shown in Table 6, the R² values of the individual stage-specific models ranged from 0.10 to 0.50. The RMSE values ranged from 0.14% to 0.21%.

Table 6.

Stage-specific baseline models.

growth stage	R2	RMSE (%)
period_1	0.10	0.19
period_2	0.35	0.17
period_3	0.28	0.21
period_4	0.49	0.14
period_5	0.50	0.21
period_6	0.50	0.20
period_7	0.40	0.21
period_8	0.27	0.20
Whole growth cycle	0.87	0.21

The aggregated whole-growth-cycle performance was relatively high. However, the R² values within individual stages were low. This indicates that independently trained stage-specific models were not sufficiently stable. This may be related to the reduced sample size within each individual growth stage. It may also be related to the limited LNC variation within each stage. Therefore, stage-specific modeling can partially reduce phenological heterogeneity. However, it may also weaken model robustness. This is because useful information cannot be shared across different stages.

3.3.3. GD-LNC without growth-stage embedding

To evaluate the contribution of growth-stage embedding, the growth-stage embedding branch was removed from the complete GD-LNC model. In this ablated configuration, the model retained the main spectral feature extraction structure. It also retained the regression structure. However, no explicit phenological information was provided. As shown in Table 7, GD-LNC without growth-stage embedding achieved an R² of 0.70 and an RMSE of 0.32% on the test set. Compared with the complete GD-LNC model, its performance clearly decreased across all datasets. In particular, the test R² decreased from 0.88 to 0.70. The test RMSE increased from 0.21% to 0.32%.

Table 7.

Component ablation of GD-LNC.

Modeling method	Training set		Validation set		Test set
	R2	RMSE (%)	R2	RMSE (%)	R2	RMSE (%)
GD-LNC without growth-stage embedding	0.74	0.30	0.75	0.28	0.70	0.32
GD-LNC without the dynamic fusion gate	0.83	0.24	0.83	0.23	0.84	0.24
GD-LNC	0.88	0.20	0.85	0.22	0.88	0.21

This decline indicates that the GD-LNC network structure alone was insufficient to achieve the best LNC estimation performance. The comparison also suggests that the improvement of the complete GD-LNC model was not only caused by increased model complexity. Instead, growth-stage embedding provided important phenological context for interpreting spectral features. It also contributed substantially to the accuracy improvement of LNC estimation.

3.3.4. GD-LNC without the dynamic fusion gate

To evaluate the contribution of the dynamic fusion gate, an ablated version of GD-LNC was constructed by replacing the gating mechanism with direct feature concatenation. In this configuration, growth-stage embedding was retained. However, spectral features and growth-stage features were combined directly. No adaptive feature recalibration was used. As shown in Table 7, GD-LNC without the dynamic fusion gate achieved an R² of 0.84 and an RMSE of 0.24% on the test set. Compared with the complete GD-LNC model, the test R² decreased from 0.88 to 0.84. The test RMSE increased from 0.21% to 0.24%.

This result indicates that direct concatenation of spectral and growth-stage features was less effective than adaptive feature fusion. The dynamic fusion gate further enhanced the model. It adaptively adjusted feature contributions under different growth-stage conditions. Therefore, it improved the overall accuracy and robustness of LNC estimation.

3.3.5. Leave-one-date-out validation

Under the leave-one-observation-date-out validation strategy, all predictions from the eight held-out observation dates were pooled to evaluate the overall predictive performance. The pooled results showed an R² of 0.27, RMSE of 0.50, indicating that the model retained a certain level of predictive ability under a stricter temporal validation strategy.

3.4. Comparison of model accuracy based on different sensitive feature sets and modeling algorithms

The sensitive spectral wavelengths selected by SPA and SFLA were each used to develop LNC estimation models with partial least squares regression (PLSR), long short-term memory (LSTM) networks, random forest regression (RF), and the growth-stage deep learning model for Leaf Nitrogen Content estimation in winter wheat (GD-LNC). This resulted in eight model combinations. These models were comparatively evaluated on both the training and independent validation sets, with results presented in Table 8, Fig. 7, Fig. 8.

Fig. 7.

LNC estimation results from different modeling methods. Note: RF, random forest; PLSR, partial least squares regression; GD-LNC, a Growth-stage Deep learning model for Leaf Nitrogen Content estimation in winter wheat; LNC, leaf nitrogen content.

Fig. 8.

LNC estimation results from Long Short-Term Memory. Note: LSTM, long short-term memory; LNC, leaf nitrogen content.

Overall, GD-LNC demonstrated the best performance across both feature sets, followed by RF and LSTM, with PLSR performing relatively weakly. Using the SPA feature set, GD-LNC achieved R² = 0.72 and RMSE = 0.30% on the test set, while RF, LSTM and PLSR reached R² = 0.62, RMSE = 0.35%; R² = 0.60, RMSE = 0.36%; and R² = 0.49, RMSE = 0.41%, respectively.

When using the SFLA feature set, the test performance of all four models was further improved, with GD-LNC achieving R² = 0.88 and RMSE = 0.21%, RF reaching R² = 0.79 and RMSE = 0.27%, LSTM reaching R² = 0.83 and RMSE = 0.24%, and PLSR attaining R² = 0.79 and RMSE = 0.27%.

From the perspective of feature set differences, SFLA not only reduced feature dimensionality but also maintained or improved model accuracy, with more stable gains observed for GD-LNC and RF. This indicates that the band combinations selected by SFLA possess stronger discriminative power and robustness under multi-temporal, cross-year grain-filling conditions. Based on the validation set results, the SFLA + GD-LNC combination emerged as the optimal approach. It provides an optimal model foundation for the continuous estimation and mapping of LNC in the subsequent grain-filling period.

3.5. Mapping leaf nitrogen content of winter wheat using GD-LNC

Based on the above model comparison results, the optimal approach in the independent validation set was the combination of SFLA-selected sensitive bands with GD-LNC. This approach achieved high accuracy on both the training and validation sets, providing a reliable model foundation for subsequent image-level inversion and mapping. Accordingly, the SFLA + GD-LNC model was applied to UAV hyperspectral imagery to perform continuous LNC estimation at key growth stages—Z71, Z73, Z75, Z83, and Z91—and to generate field-scale spatial distribution maps of LNC, as illustrated in Fig. 9.

Fig. 9.

Field-scale spatial distribution of winter wheat LNC based on GD-LNC.

Based on the estimation results, the LNC distribution maps generated by SFLA + GD-LNC clearly captured the dynamic changes of LNC across different periods and accurately restored the spatial differences among treatment plots. Most fields exhibited small errors and consistent variation trends.

4. Discussion

4.1. Characterization differences of sensitive bands between SPA and SFLA

Although SPA and SFLA selected different specific wavelengths, they share a consistent “nitrogen-nutrition orientation” in terms of spectral functionality. They cover regions from the visible light to the red-edge-near-infrared that are closely related with chlorophyll absorption and canopy structure. These spectral regions can comprehensively characterize plant nitrogen status by integrating information on pigment content, leaf structure, and canopy growth. During the grain-filling period, LNC typically decreases as nitrogen is translocated to the grains, accompanied by reductions in chlorophyll content and changes in photosynthetic capacity. Consequently, visible-light absorption features and the position and shape of the red edge are sensitive to nitrogen variations [30]. The differences between the two methods are mainly reflected in the “clustering” and “target orientation” of the selected band combinations. SFLA forms more concentrated band clusters in the blue-green zone and the red-light zone. It retains key red-edge and near-infrared windows. This configuration is more conducive to capturing the joint constraints of pigment changes and canopy structure variations on LNC during the grain-filling period. In contrast, SPA selects relatively discrete bands, likely prioritizing the reduction of multicollinearity rather than directly optimizing for prediction error [47]. Considering the drift of the spectral-physiological relationship in different observation phases during the grain-filling period. It is difficult for a single-phase training model to migrate stably. [31]. SFLA performs global search and local updates with the objective of minimizing inversion error. This results in band combinations that better capture the primary response regions of LNC while controlling redundancy. This often leads to greater stability and overall accuracy in multi-temporal modeling [40,41].

4.2. Growth-stage awareness and global modeling advantages of GD-LNC

Compared with PLSR and RF, the advantage of GD-LNC lies in its ability to address two key challenges in multi-temporal inversion during the grain-filling stage: (i) the presence of long-range dependencies and complex correlations among hyperspectral bands, and (ii) the filling process causes the correspondence between LNC and spectral characteristics to undergo stage-dependent drift as the phase changes [30]. Traditional convolutional structures are more inclined toward local pattern extraction and struggle to adequately represent global inter-spectral dependencies. Moreover, cascaded convolution may also introduce repetitive feature extraction. This leads to computational redundancy and reduces efficiency and generalization performance in multi-temporal tasks [36]. GD-LNC introduces the Transformer self-attention mechanism in a 1D CNN-based spectral encoder. This is conducive to explicitly modeling the global correlation across bands. It enhances the information expression of key spectral bands [24]. It also retains the strengths of convolutional networks in hierarchically extracting discriminative features [23]. More importantly, the model integrates observation temporal and growth-stage information[48]. This enables the network to learn spectral features while incorporating phenological characteristics. This design introduces a dynamic weighting mechanism. Thereby, it enhances the stability and generalization capability of LNC inversion across time. The explicit growth-stage embedding and the dynamic fusion gating mechanism are complementary rather than redundant. Growth-stage embedding provides a static phenological conditioning signal, indicating which phase the observation belongs to, but not how much weight different features should receive. Conversely, a fusion gate without explicit stage conditioning must infer phenological context from spectral data alone. This process is unreliable because adjacent growth stages often exhibit overlapping spectral signatures. By coupling both mechanisms, GD-LNC achieves informed adaptivity. The embedding anchors spectral interpretation within a phenological framework. Meanwhile, the fusion gate adaptively adjusts feature contributions. Together, these mechanisms improve cross-stage stability. This is consistent with previous findings on the effectiveness of deep learning for LNC estimation [25].

4.3. Model application prospects and improvement strategies

This study innovatively developed a growth-stage deep learning model for UAV-based hyperspectral estimation of leaf nitrogen content in winter wheat. By adopting a hierarchical architecture comprising source feature extraction, dynamic deep feature fusion, and regression prediction, the model effectively enhances the consistency of multi-source feature representations across different growth stages. It also improves the complementarity of these feature representations across stages. At the application level, UAV-based hyperspectral estimation of LNC can support nitrogen diagnosis and management decision-making during the grain-filling period. It provides data support for optimizing topdressing practices and improving nitrogen use efficiency [49]. Owing to their high spatiotemporal resolution and rapid deployment advantages, UAV platforms are well suited for multi-temporal monitoring at both plot and field scales [11,12]. It should be noted that a small number of plots in the LNC mapping still exhibited some degree of systematic underestimation. As shown in Fig. 8, systematic underestimation was observed during the Z71 period in plots situated in the first six rows and the 6th to 8th columns from the left. The predicted LNC values were approximately 0.08%–0.57% lower than the field measurements. This may be attributed to a combination of factors: (1) insufficient local canopy coverage, where exposed soil background introduces non-photosynthetic spectral signals that dilute the chlorophyll-related reflectance features; (2) variations in illumination conditions caused by differences in solar zenith angle and shadow fraction across multi-temporal acquisitions; and (3) micro-environmental heterogeneity within plots, such as mixed understory vegetation or localized moisture gradients. Such non-nitrogen-related disturbances can amplify errors in multi-temporal applications [31]. A further examination of the corresponding original high-resolution imagery confirmed that the affected plots did exhibit significantly lower canopy coverage. The vegetation coverage in these plots was approximately 30%–35% lower than the average level across all plots. This condition likely contributed to the observed underestimation through increased soil background interference. To quantify canopy coverage, the original high-resolution UAV imagery was clipped according to the plot boundaries. Vegetation pixels were then separated from soil and background pixels using image-based vegetation segmentation. Canopy coverage was calculated as the proportion of vegetation pixels relative to the total number of pixels within each plot. This followed the commonly used vegetation-fraction calculation method in UAV-based crop monitoring studies [50,51]. Future improvements will focus on multi-source data fusion and consistency control. The advantages of hyperspectral data will be retained. Auxiliary information representing canopy structure and growth vigor will be incorporated to enhance the model's ability to distinguish background and structural differences. Coupled with cross-date radiometric consistency and normalization, this approach can reduce the impact of illumination variations. Additionally, further optimization of growth-stage encoding and sample coverage can improve the model's robustness and transferability on anomalous plots [25,30].

For practical UAV-based LNC monitoring, the use of growth-stage information also requires consideration of intermediate or unseen phenological stages. For intermediate stages within the grain-filling period, the proposed high-frequency monitoring strategy provides a practical way to reduce uncertainty. Samples located between two observed stages can be assigned to the closest Zadoks stage. They can also be assigned to the corresponding phenological interval based on field observation. Because the growth-stage label reflects physiological development rather than calendar date, this strategy helps reduce phenological mismatch. This mismatch may be caused by regional, annual, or management-related differences in crop development. Thus, the growth-stage embedding provides a physiological reference for aligning samples across different dates and regions. It supports more stable LNC estimation within the grain-filling period.

Nevertheless, the current GD-LNC model should not be regarded as an unrestricted extrapolation model for all possible unseen growth stages. Its cross-stage adaptation refers to improved robustness across the observed and physiologically adjacent grain-filling stages. For growth stages far outside the training range, additional samples and further validation are still required.

Although the proposed GD-LNC model showed satisfactory inversion performance under the current experimental conditions, several aspects still deserve further investigation. This study mainly focused on the changes of LNC at multiple phases during the grain-filling stage under the conditions of multiple varieties and various water and fertilizer supplies. Further studies involving more diverse ecological regions, wheat varieties and environmental conditions may help enhance the universality, robustness, and applicability of the framework. Future work may also incorporate multi-source information and cross-date consistency strategies to improve model stability under complex field conditions further.

5. Conclusions

This study carried out the field experiments of winter wheat in two growing seasons. It established an integrated estimation workflow of “sensitive band selection, growth-stage fusion, and image-level LNC inversion and mapping.” This workflow enables growth-stage monitoring and mapping of leaf nitrogen content based on UAV hyperspectral imagery. The core innovation of GD-LNC lies in the synergistic integration of explicit growth-stage embedding with a dynamic feature fusion gating mechanism. This coupled design enables the model to simultaneously account for phenological context and adaptively recalibrate spectral feature contributions across temporal observations. The results indicate that both SPA and SFLA effectively selected spectral bands significantly correlated with LNC. However, the band combinations selected by SFLA were more concerntrated in the blue-green and red band range. They also retained key red-edge to near-infrared information. This resulted in higher overall accuracy and greater stability for models based on the SFLA feature set. In LNC estimation, GD-LNC outperformed both RF, LSTM and PLSR across both feature sets, with the SFLA + GD-LNC combination achieving the best performance: R² = 0.88 and RMSE = 0.20% on the training set, and R² = 0.85 and RMSE = 0.22% on the validation set, and R² = 0.88 and RMSE = 0.21% on the test set. The LNC spatial distribution maps for key periods of the grain-filling stage, generated based on the optimal approach, clearly capture the dynamic changes of LNC across different dates. In summary, the GD-LNC proposed in this study provides a feasible approach for UAV-based hyperspectral estimation of nitrogen in winter wheat. Dense phenological monitoring during the grain-filling period can reduce stage-mismatch effects. These effects may occur across different dates, years, and regions. However, the applicability of GD-LNC to growth stages outside the observed grain-filling range still requires further validation.

Authorship contribution statement

Siwen Zhao: Writing - original draft, Visualization, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptu-alization. Chao Song: Writing - review & editing, Validation, Su-pervision, Methodology, Formal analysis. Baoyuan Zhang: Visualization, Software, Methodology, Formal analysis, Data curation. Xingyu Liu: Visualization, Software, Methodology, Investigation, Datacuration. Fajian Zong: Visualization, Software, Methodology, Formalanalysis. Shengnan Zhai: Software, Methodology, Investigation, Datacuration. Xiaojun Liu: Resources, Investigation, Formal analysis, Datacuration. Liping Chen: Methodology, Formal analysis. Xiaohe Gu: Writing - review & editing, Validation, Supervi-sion, Resources, Project administration, Funding acqui-sition.

Data availability

Data will be made available on request. The data that support the findings of this study are available from the corresponding author, guxh@nercita.org.cn (X.G.), upon reasonable request. The trained model weights is publicly available in the GitHub repository at https://github.com/zsww1214-ai/model-weight.

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.