KG-FT-Transformer: a knowledge-guided feature tokenizer transformer for hyperspectral prediction of lettuce quality trait and fingerprint analysis

Abstract

Accurate and non-destructive prediction of lettuce quality traits is essential for variety identification, germplasm utilization, and intelligent breeding. However, existing approaches relying on handcrafted features or purely data-driven models face limitations under small-sample conditions, including constrained prediction accuracy, weak interpretability, and an increased risk of overfitting. To address these challenges, we propose a knowledge-guided feature tokenizer transformer (KG-FT-Transformer) framework for hyperspectral quality trait prediction and fingerprint analysis. This framework integrates domain prior knowledge with data-driven learning, significantly improving prediction accuracy while enhancing biological interpretability. The KG-FT-Transformer employs a Transformer-based architecture integrating multi-head attention (MHA) with a gated feed-forward network (GFFN), enabling nonlinear spectral modeling and rich feature interactions. We evaluated its performance on three key quality traits: relative chlorophyll content (SPAD), soluble solids content (SSC), and moisture content (MC). The model achieved R² values of 0.9534, 0.9185, and 0.9226, with corresponding residual predictive deviation (RPD) values of 4.63, 3.50, and 3.60, outperforming all baseline models and demonstrating stable and consistent prediction performance. Moreover, pixel-wise predictions were used to construct quality trait fingerprints through pseudo-color mapping, intuitively visualizing the spatial distribution and varietal specificity of traits. SPAD and SSC exhibited visually consistent central aggregation patterns, while MC revealed distinct spatial variations among cultivars. These fingerprint-based representations provide spatially informed, qualitative references that may assist the interpretation of DUS-related (Distinctness, Uniformity, Stability) trait characteristics. Overall, this study demonstrates the potential of integrating hyperspectral prediction with quality fingerprinting for non-destructive quality assessment and breeding-oriented analysis, and provides a complementary perspective for germplasm identification and crop improvement.

Highlights

• •KG-FT-Transformer is proposed for lettuce quality trait prediction under small-sample conditions.
• •The model accurately and stably predicts SPAD, SSC, and MC for rapid, non-destructive quality assessment.
• •Pixel-wise maps serve as quality fingerprints for spatial visualization, DUS interpretation, and germplasm analysis.

1. Introduction

Lettuce (Lactuca sativa L.) is one of the world's most important leafy vegetables. With its short growth cycle, broad adaptability, high nutritional value, and diverse consumption patterns, it plays a vital role in ensuring vegetable supply and improving human diets (Dyląg et al., 2023). In the context of growing concerns over food security and sustainable agriculture, research on lettuce production and quality has become increasingly critical (Qin et al., 2025). Commercial lettuce cultivars are generally classified into several major horticultural groups, including Butter (B), Crisphead (C), Leaf (L), Oakleaf (O), Romaine (R), Stem (S), and Wild relatives (W). These groups differ in traits such as leaf morphology, color, and heading characteristics, and collectively represent the broad phenotypic and genetic diversity present in lettuce germplasm (Chen, Chen, et al., 2025; Guo et al., 2023). This diversity provides an important biological foundation for systematic quality evaluation and comparative analysis across cultivars.

Beyond yield and resistance (Lima et al., 2023; Urbutis et al., 2024), lettuce quality traits have become a primary criterion for cultivar evaluation (Lian et al., 2021). These traits determine consumer perception of taste, nutrition, and freshness, and directly influence marketability. Lettuce quality is commonly described by physicochemical parameters such as relative chlorophyll content (SPAD) (Mendoza-Tafolla et al., 2019; Putti et al., 2025), soluble solids content (SSC) (Eshkabilov et al., 2021; Zhao & Xie, 2025), fresh weight (FW), dry weight (DW), and moisture content (MC) (Bu et al., 2025). Among them, SPAD serves as an indirect indicator of nitrogen status and photosynthetic capacity (Putti et al., 2025; Chen, Huang, et al., 2025). SSC reflects sugar concentration and soluble compounds, shaping flavor and sweetness (Liu et al., 2025; Yu et al., 2022). MC relates to leaf crispness and consumer acceptance (Alemu & Oanh, 2025; Ji et al., 2025). Together, these traits represent essential indicators of consumer preference and practical standards for lettuce quality assessment.

Traditional physicochemical assays, though highly accurate (Avelar et al., 2025; Baron et al., 2024; Hao et al., 2023), are often time-consuming, labor-intensive, and destructive. These drawbacks limit their suitability for large-scale, rapid, and non-destructive monitoring required in modern agricultural and food production systems (Barbosa et al., 2023; Rodionova et al., 2024). To address this, spectral analysis has been increasingly applied for crop quality assessment. By capturing reflectance or absorbance at specific wavelengths, spectral signatures can indirectly reveal internal quality attributes of crops. This approach has been successfully used for non-destructive estimation of chlorophyll content, nutrient composition, heavy metal accumulation, and various physicochemical traits. For example, (Eshkabilov et al., 2022) explored the dynamic effects of N, P, and K compounds on lettuce growth using hyperspectral imaging (HIS). Through principal component analysis (PCA), partial least squares regression (PLSR), and variable importance projection (VIP) analysis, they identified six key bands (390.57–438.02 nm, 497–550 nm, 551–600 nm, 681.34–774 nm, 802–821 nm, and 822–838 nm) for modeling. Their results achieved R² values above 0.85, enabling accurate characterization of nutrient dynamics in lettuce leaves. Similarly, (Li et al., 2024) focused on quality parameters during the rosette stage. By combining spectral preprocessing with feature band selection, they established predictive models for SSC, vitamin C, and MC using SG-UVE-PLSR and Nor-CARS-PLSR. The models achieved R² values of 0.8648, 0.8578, and 0.8047, respectively, and were integrated into a real-time software platform for quality monitoring. However, chemometric approaches such as PLSR and PCA (Falcioni et al., 2022; George et al., 2024; Zhou et al., 2018) typically rely on complex preprocessing (e.g., denoising, normalization, spectral correction) and manual band selection. These steps increase modeling complexity and often limit generalizability across environments.

To overcome these limitations, machine learning (ML) methods have been increasingly applied in spectral modeling. Algorithms such as support vector regression (SVR), random forest (RF), and gradient boosting decision trees (GBDT) (Lu et al., 2025; Mokhtar et al., 2022; Zhou et al., 2024) show strong capabilities in handling nonlinear relationships and high-dimensional data, thereby reducing the need for manual intervention. For instance, (Sun et al., 2026) developed an Support Vector Machine (SVM) model by combining HIS with heterogeneous two-dimensional spectral correlation analysis, enabling non-destructive prediction of cadmium (Cd) content in lettuce. The model achieved an R² of 0.8543, offering a reliable solution for Cd monitoring. Similarly, (Gonçalves et al., 2025) evaluated seven machine learning algorithms for predicting chlorophyll and pigment compounds in lettuce leaves. Their results demonstrated that PLSR and RF were the most robust approaches for pigment estimation. Using RF, they achieved R² values of 0.91, 0.76, and 0.79 for anthocyanins, flavonoids, and phenolic compounds, respectively. These studies provide a non-destructive framework for biochemical monitoring and quality assessment in lettuce. Despite their success, conventional machine learning models still rely on feature selection and often struggle with generalizability in complex agricultural contexts.

In recent years, deep learning (DL) has demonstrated unique advantages in spectral modeling. Convolutional neural networks (CNN) and attention-based neural networks (ACNN) can directly extract features from high-dimensional raw spectra, effectively capturing complex nonlinear relationships. This capability eliminates the need for extensive preprocessing and manual feature engineering, thereby improving both modeling efficiency and predictive accuracy (Furbank et al., 2021; Yu et al., 2022). Despite these advantages, many existing deep learning-based spectral models primarily rely on direct mappings from spectral inputs to target variables, and their performance can become unstable under limited sample conditions, where overfitting and reduced prediction consistency may arise. In practical crop quality assessment, spectral-trait relationships are often influenced by varietal characteristics, which can introduce additional variability that is difficult to robustly capture when training data are limited. In addition, while recent studies have begun to explore spatial prediction, most deep learning approaches still primarily focus on optimizing global prediction accuracy, and the spatial expression of predicted traits is often considered supplementary rather than integral to the modeling process.

These considerations indicate that further advances in hyperspectral quality prediction, particularly under limited sample conditions, require a representation learning strategy that not only exploits the expressive capacity of deep neural networks, but also introduces appropriate structural guidance to improve learning stability and prediction consistency. Allowing structured prior information to participate directly in feature interaction provides a principled means to constrain model optimization and to better distinguish intra- and inter-varietal spectral patterns that are difficult to infer from data alone. In this context, incorporating varietal category knowledge into the representation space provides an effective way to guide feature learning and reduce ambiguity arising from heterogeneous spectral-trait relationships. Meanwhile, a token-based formulation allows spectral features and prior knowledge to be jointly modeled within a unified interaction framework, and naturally extends conventional sample-level regression toward pixel-wise inference, thereby linking numerical prediction with spatial trait expression.

In recent years, Transformer models have shown strong capability in modeling complex feature dependencies through self-attention mechanisms and have been successfully applied in domains such as computer vision and natural language processing. However, most existing Transformer architectures are primarily designed for specific data modalities and are not directly optimized for spectral regression tasks, particularly in small-sample settings. To address this gap, it is necessary to adapt and tailor Transformer-based frameworks to the characteristics of hyperspectral data, enabling efficient modeling of high-dimensional spectral variations while maintaining stable optimization behavior. Accordingly, this study integrates hyperspectral data and varietal category knowledge within a Transformer-based architecture, with the core objective of guiding feature interaction through structured prior information. By embedding category knowledge into the learning process, the model is encouraged to align spectral representations with varietal context, which helps improve training stability and predictive accuracy under limited sample conditions.

To this end, we propose a knowledge-guided feature tokenizer transformer (KG-FT-Transformer) for hyperspectral prediction of lettuce quality traits. The proposed model embeds varietal category knowledge into spectral feature interactions, enabling joint modeling of spectral reflectance patterns and varietal relationships. By integrating multi-head attention (MHA) with a gated feed-forward network (GFFN), the framework enhances nonlinear modeling capacity and feature interaction efficiency. Using SPAD, SSC, and MC as representative quality traits, the proposed approach achieves accurate prediction and further generates pixel-wise pseudo-color fingerprint maps, which provide intuitive visualization of spatial trait distribution and support exploratory interpretation related to quality differentiation and DUS-relevant characteristics.

The main contributions of this study are summarized as follows:

• (1)A knowledge-guided feature tokenizer Transformer is developed by embedding varietal category knowledge and spectral reflectance information into a unified interaction framework, enabling effective modeling of intra- and inter-varietal spectral relationships for high-precision quality trait prediction.
• (2)The proposed approach achieves accurate and stable prediction of SPAD, SSC, and MC, supporting rapid and non-destructive assessment of lettuce quality traits.
• (3)Pixel-wise predictions are visualized as quality trait fingerprints, offering an intuitive spatial representation of predicted traits and providing complementary references for quality visualization, exploratory DUS-related interpretation, and germplasm analysis.

2. Materials and methods

2.1. Lettuce cultivation

To ensure a representative evaluation of quality trait prediction, we selected 150 major commercial cultivars representing seven subgroups: Butter (B), Crisphead (C), Leaf (L), Oakleaf (O), Romaine (R), Stem (S), and Wild relatives (W). All lettuce cultivars were provided by the Institute of Biotechnology, Beijing Academy of Agriculture and Forestry Sciences (BAAFS). These cultivars have been previously subjected to whole-genome sequencing and genotyping analyses, exhibiting clear genetic differentiation and trait-level diversity, which supports their suitability as a genetically diverse population for quality trait evaluation. The experiment was conducted on March 18, 2023, in a controlled greenhouse at the BAAFS (39.9438°N, 116.2876°E). Prior to sowing, seeds were surface-sterilized and germinated in 72-cell seedling trays under standard nursery management. At the “two-leaf, one-heart” stage, uniform and healthy seedlings were transplanted into standard plastic pots (32 cm diameter, 34 cm height). The growth substrate consisted of sand, compost, and nutrient soil mixed at a 4:2:4 volume ratio, with nutrient levels and pH adjusted uniformly to ensure consistent growth conditions. Greenhouse parameters were regulated by an intelligent control system: daytime and nighttime temperatures were maintained at 25–18 °C, relative humidity was kept between 75 and 95%, and photoperiod was set to 12 h per day (06:00–18:00). A standardized fertigation regime was applied throughout the growth cycle using a drip irrigation system supplying a commercial nutrient solution (DEWODUO, Hebei Dewoduo Fertilizer Co., Ltd., China). The pH of the nutrient solution was maintained at approximately 5.5–6.5, and the electrical conductivity (EC) was controlled at about 1.2–2.5 mS·cm⁻¹. Light was mainly provided by natural greenhouse illumination, with average light intensity maintained at approximately 200–400 μmol·m⁻²·s⁻¹ during the photoperiod. At 45 days after sowing, when plants reached commercial maturity, three uniform plants per cultivar were sampled for quality trait measurements and hyperspectral image acquisition. The lettuce planting scene is shown in Fig. 1. This cultivation and sampling design was intended to ensure reasonable representativeness of quality trait variation among lettuce cultivars under controlled conditions. By selecting major commercial cultivars spanning diverse genetic backgrounds and applying uniform management practices, the dataset captures both intra-cultivar and inter-cultivar variability while minimizing environmental and acquisition-related confounding factors. Accordingly, the resulting dataset provides a consistent experimental basis for quality trait modeling and fingerprint-based trait characterization in this study.

Fig. 1.

Lettuce cultivation under controlled greenhouse conditions.

2.2. Quality trait measurement of lettuce

2.2.1. SPAD measurement

For each plant, three fully expanded, lesion-free functional leaves from the top canopy were selected. Leaf chlorophyll content was measured at the central region of each leaf using a portable chlorophyll meter (SPAD-502 Plus, Konica Minolta, Japan). To minimize bias from local light heterogeneity, each leaf was measured three times and averaged. The mean SPAD value of the three leaves was used as the plant-level trait. Measurements were conducted under uniform illumination, with leaf surfaces kept clean and moisture-free to ensure data stability and reproducibility.

2.2.2. SSC measurement

The outermost fully expanded functional leaf was chosen for SSC analysis, ensuring both structural representativeness and consistency with spectral data. Immediately after collection, leaves were chopped and juice extracted using a manual juicer (LKM-ZZ01, Likemei Tech., China). Undiluted juice was directly placed on a digital refractometer (PAL-1, ATAGO, Japan) to record Brix%. Each sample was measured three times, with the prism rinsed and dried between replicates. The average value was used as the final SSC result. All operations followed ISO 2173:2003 standards, and instruments were thoroughly cleaned before and after use to prevent cross-contamination.

2.2.3. MC measurement

MC was determined following AOAC (1990) guidelines after hyperspectral imaging. FW was recorded after gently wiping residual surface water with absorbent paper. Samples were then placed in pre-labeled aluminum boxes and oven-dried at 105 °C for 24 h. Dried samples were cooled in a desiccator for 30 min and weighed to obtain DW. MC was calculated as:

$$M=\frac{m_0-m_1}{m_0}$$ (1)

where ${\mathrm{m}}_0$ is FW and ${\mathrm{m}}_1$ is DW. Results were expressed as percentages. All samples were independently processed, and each measurement was repeated three times to ensure accuracy and reproducibility.

2.3. Hyperspectral imaging acquisition and spectral extraction

2.3.1. Hyperspectral imaging acquisition

The HSI system operated in a visible–near infrared (Vis-NIR) reflectance mode. All imaging experiments were conducted in a fully enclosed darkroom (2.5 m × 2.4 m × 3.4 m) to eliminate ambient light interference and ensure high-quality image acquisition. The system comprised four major components: (1) a push-broom hyperspectral imager (GaiaField-V10E, Dualix Spectral Imaging, China); (2) a high-resolution CCD industrial camera equipped with a 23 mm standard lens; (3) four 50 W halogen‑tungsten lamps providing continuous-spectrum illumination; and (4) a computer workstation (Dell Ins 15-7510-R1645S, USA) running SpecVIEW software for data acquisition and control. The spectrometer covered a wavelength range of 366.36-976.11 nm with 256 contiguous bands and a spectral resolution of 2.8 nm, enabling fine capture of the Vis-NIR spectral responses of lettuce samples. The camera provided a spatial resolution of 696 pixels per line using a line-scan imaging mode. The illumination system delivered a broad continuous spectrum (350–2500 nm), effectively simulating natural light and ensuring uniform illumination. During imaging, the lens-to-sample distance was fixed at 2.5 m. Lamps were symmetrically arranged at a 45° angle relative to the optical axis to optimize light uniformity and minimize specular reflection. Acquisition parameters were set as follows: exposure time 30 ms, frame rate 14 FPS, gain 1. Spatial binning of 2 and spectral binning of 4 were applied to balance the signal-to-noise ratio and data acquisition efficiency. The final hyperspectral images had a spatial-spectral resolution of 696 × 775 × 256 (lines × columns × bands), providing complete spectral-spatial information for each sample.

2.3.2. Spectral data extraction

The acquired hyperspectral images were subjected to a series of standardized preprocessing steps to obtain representative spectral features for subsequent modeling. First, black-white reference correction was performed to eliminate spectral deviations caused by uneven illumination and environmental interference. A 99% reflectance polyvinyl chloride (PVC) panel was scanned as the white reference to calibrate illumination variation. Dark current images were collected by switching off the lamps and covering the camera lens. These reference images were used to compensate for the inherent dark current of the hyperspectral camera. The reflectance spectra of lettuce samples were then calculated using the following standard transformation formula:

$$R=\frac{R_{\mathit{raw}}-R_d}{R_w-{\mathrm{R}}_d}$$ (2)

where R is the calibrated reflectance, $R_{\mathit{raw}}$ is the raw reflectance, $R_d$ is the dark reference, and $R_w$ is the white reference.

After image calibration, a high-throughput hyperspectral data processing workflow was established, as shown in Fig. 2. First, a representative single-band image was selected from the three-dimensional hyperspectral data cube. Band 210 was chosen in this study due to its high signal-to-noise ratio and stable illumination response, which enhanced the contrast between lettuce canopy and background. The selected band was processed using the OTSU adaptive thresholding method, which automatically determined the optimal threshold to segment the foreground (lettuce canopy) from the background. This produced a binary mask image. The mask was further refined through connected component analysis in OpenCV, where noise regions smaller than 150 pixels were removed, and the main canopy region was retained as the region of interest (ROI). The ROI mask was then applied to the entire hyperspectral cube. Pixel-wise reflectance values across all 256 bands were extracted, with background pixels excluded. Finally, the average spectral reflectance curve within the ROI was computed, resulting in a 256-dimensional feature vector for each lettuce sample. This workflow ensured standardization and consistency in spectral data extraction, while balancing high throughput with accuracy. It provided a robust foundation for subsequent regression modeling of quality traits.

Fig. 2.

Workflow of lettuce hyperspectral images acquisition, ROI determination, and spectral data extraction.

2.4. Modeling methods

2.4.1. Knowledge-guided feature tokenizer transformer

Hyperspectral reflectance data provide rich biochemical and structural information for crop quality assessment, but their high dimensionality and strong inter-band correlations pose challenges for effective representation learning. In addition, lettuce quality traits are influenced by varietal differences, such that samples from different cultivars may exhibit distinct spectral-trait relationships even when acquired under identical conditions. Under limited sample conditions, this varietal heterogeneity further increases the difficulty of learning stable and discriminative spectral representations. Existing spectral deep learning models predominantly learn direct mappings from spectral inputs to target variables in a purely data-driven manner, which limits their ability to explicitly encode cultivar-related structure and to differentiate intra- and inter-varietal spectral patterns, particularly when training data are scarce.

To enhance representation learning for high-dimensional spectral features and to explicitly introduce structured varietal knowledge, we developed a knowledge-guided feature tokenizer transformer (KG-FT-Transformer), inspired by (Gorishniy et al., 2021). The proposed model is designed to jointly model hyperspectral reflectance features and varietal category information within a unified interaction framework, enabling regression prediction of lettuce quality traits from hyperspectral data. It consists of two main modules: feature embedding and feature learning with prediction, as shown in Fig. 3. In the embedding stage, both variety category knowledge and hyperspectral reflectance features were embedded using a feature tokenizer. In the learning stage, the embedded features were concatenated and fed into a transformer encoder for feature representation learning, followed by a linear regression layer for prediction. Here, the concatenation is performed at the token level to form a token sequence, rather than at the feature level to construct a single augmented feature vector. Considering the diversity of lettuce varieties, variety category knowledge nodes were explicitly introduced during feature embedding to improve the model's ability to capture intra- and inter-varietal relationships. Specifically, each sample (e.g., B222–1, C545–1) was assigned a categorical varietal label (e.g., 222, 545), which was subsequently mapped to a learnable knowledge token via the feature tokenizer. The feature tokenizer then transformed the categorical knowledge feature $x^{\left(\mathit{cat}\right)}$ into a knowledge node embedding vector $T_{\mathit{embed}}^{\left(\mathit{cat}\right)}\in R^{k\times d}$. The embedding representation of the variety category knowledge node can be expressed as:

$$T_i^{\left(\mathit{cat}\right)}=x_i^{\left(\mathit{cat}\right)}\bullet W_i^{\left(\mathit{cat}\right)}+b_i^{\left(\mathit{cat}\right)}\in R^d$$ (3)

where $b_i^{\left(\mathit{cat}\right)}$ is the i-th feature bias, and $W_i^{\left(\mathit{cat}\right)}$ is the weight matrix. In this study, the varietal category is represented by a single knowledge token ($k$ = 1) for each sample, which encodes prior varietal information in a learnable latent space. Similarly, the embedding of the spectral feature sequence nodes is defined as:

$$T_i^{\left(\mathit{band}\right)}=x_i^{\left(\mathit{band}\right)}\bullet W_i^{\left(\mathit{band}\right)}+b_i^{\left(\mathit{band}\right)}\in R^d$$ (4)

where $x_i^{\left(\mathit{band}\right)}$ denotes the reflectance value of the i-th spectral band, and $W_i^{\left(\mathit{band}\right)}$ and $b_i^{\left(\mathit{band}\right)}$ are learnable parameters. This formulation maps each spectral band into the same latent space as the knowledge token, enabling unified modeling of heterogeneous inputs.

Fig. 3.

Architecture of the KG-FT-Transformer model.

Subsequently, the variety category knowledge embedding and the spectral embedding were stacked together to construct an integrated feature representation that simultaneously incorporates prior variety knowledge and spectral information. This stacked embedding can be expressed as:

$$T=\text{stack}\left[T_1^{\left(\mathit{cat}\right)},\cdots ,T_k^{\left(\mathit{cat}\right)},T_1^{\left(\mathit{band}\right)},\cdots ,T_m^{\left(\mathit{band}\right)}\right]\in R^{\left(k+m\right)\times d}$$ (5)

This stacked representation forms a sequence of tokens, in which the knowledge token and spectral tokens are treated as parallel inputs to the Transformer encoder. As a result, varietal category information is not injected as a fixed auxiliary feature but participates in feature interaction modeling through the subsequent self-attention mechanism.

In the feature learning stage, inspired by (Devlin et al., 2019), the embedding of the [CLS] token is appended to $T$. The resulting sequence was then processed through a series of Transformer layers $F_1,\cdots ,F_L$ to capture complex feature dependencies and hierarchical representations. Formally, the feature extraction process can be expressed as:

$$T_0=\text{stack}\left[\left[\mathit{CLS}\right],T\right],T_i=F_i\left(T_{i-1}\right)$$ (6)

Each Transformer layer consisted of a MHA module and a GFFN. The GFFN enhanced model performance by introducing dynamic feature gating and efficient parameter utilization without increasing computational overhead. Through the MHA mechanism, the knowledge token and spectral tokens can interact repeatedly across layers, allowing varietal context to modulate spectral feature learning in a data-driven manner. To further improve training stability and optimization efficiency, residual connections and PreNorm operations were applied in every layer (Wang et al., 2019).

The final representation of the $\left[\mathit{CLS}\right]$ token was utilized as the global feature descriptor and passed through a linear regression layer to generate the quality trait prediction.

$$\widehat{y}=\text{Linear}\left(\mathit{ReLU}\left(\mathit{LayerNorm}\left(T_L^{\left[\mathit{CLS}\right]}\right)\right)\right)$$ (7)

2.4.2. Model implementation and training details

2.4.2.1. Model architecture

The KG-FT-Transformer was implemented by extending the FT-Transformer architecture. The model consists of a stack of transformer encoder layers with explicit handling of categorical and continuous inputs. Specifically, the Transformer encoder comprises four layers (L = 4), each employing MHA with 16 attention heads. The hidden embedding dimension of all tokens is set to 32 ($d$ = 32). A single categorical feature corresponding to the lettuce varietal label is used and embedded as a knowledge token, while the remaining 256 spectral bands are treated as continuous tokens. The GFFN within each transformer layer follows a gated linear unit structure to enhance nonlinearity and feature interaction capacity. A learnable [CLS] token is prepended to the token sequence and is initialized randomly as a trainable parameter. The final prediction is obtained from the [CLS] token via a linear regression head with output dimension 1. Dropout regularization is applied within the transformer layers, with an attention dropout rate of 0.2 and a feed-forward dropout rate of 0.1 to mitigate overfitting.

2.4.2.2. Training strategy

Model training was performed using the mean squared error (MSE) loss function. Optimization was carried out using the AdamW optimizer, with an initial learning rate of 5 × 10⁻⁴ and a weight decay coefficient of 0.01. The model was trained with a batch size of 32 for a maximum of 2000 epochs, and model parameters were optimized using backpropagation. To reduce the risk of overfitting under limited sample conditions and to ensure stable model selection, an early stopping strategy based on calibration loss was employed. During training, model performance was monitored on the calibration set at each epoch. Training was terminated if the calibration loss did not improve for 100 consecutive epochs. The model checkpoint corresponding to the lowest calibration loss was retained as the final model.

2.4.2.3. Data splitting

The dataset consists of 450 lettuce samples, each containing one categorical varietal label and 256 hyperspectral reflectance features. The samples were randomly divided at the sample (plant) level, independent of cultivar identity, to evaluate general regression performance rather than cultivar-level generalization. Specifically, the dataset was split into a training set (60%), a calibration set (20%), and a prediction set (20%), corresponding to 270 training samples, 90 calibration samples, and 90 prediction samples, respectively. The splitting process was performed using a fixed random seed (seed = 123) to ensure reproducibility.

2.4.3. Baseline model

In previous studies on hyperspectral-based crop quality prediction, PLSR and SVR have been widely adopted due to their robustness, adaptability, and interpretability. They have achieved promising results in predicting fruit soluble solids content (Liu et al., 2025; Xiao et al., 2025), leaf nitrogen concentration (Wang et al., 2025), and water content (Wu et al., 2025; Yuan et al., 2025). Accordingly, these two classical methods were selected as baseline models for benchmarking against the proposed deep learning approach. PLSR is a well-established multivariate modeling technique, particularly effective for high-dimensional datasets with strong multicollinearity. Its principle is to maximize covariance between the spectral data matrix X and the response variable matrix Y, thereby extracting latent variables (LVs) with the strongest explanatory power for regression modeling. SVR, derived from SVM theory, is a supervised learning method designed for nonlinear regression. By mapping low-dimensional features into a high-dimensional space via kernel functions, SVR constructs an optimal regression hyperplane with enhanced nonlinear modeling capacity. To improve model generalization and predictive accuracy, grid search optimization (Song et al., 2024) was employed to determine the best parameter combinations. The search space covered dataset partitioning strategies (KS, SPXY, Random), spectral preprocessing methods (None, MA, SG, MSC, D1, D2, DT, WAVE), and feature selection techniques (None, Cars, Lars, Uve, Spa, GA, PCA).

In addition to conventional methods, several deep learning models were introduced as baseline approaches to systematically evaluate the effectiveness of the proposed KG-FT-Transformer framework. An ACNN was first adopted as a representative deep learning baseline for hyperspectral regression. By incorporating channel-level attention mechanisms into one-dimensional convolutional layers, the ACNN enhances sensitivity to informative spectral regions and improves its ability to extract nonlinear representations from raw spectral reflectance data. Furthermore, a Spectral-ViT model was included as a Transformer-based baseline to explore the applicability of vision Transformer-inspired architectures for spectral feature modeling. This model leverages self-attention mechanisms to capture global dependencies among spectral features, enabling the modeling of complex feature interactions. In addition, a hybrid CNN-Transformer architecture was constructed as a comparative model, combining the local feature extraction capability of convolutional neural networks with the global dependency modeling of Transformer encoders. In this model, convolutional modules are first used to extract low-level spectral feature representations, which are subsequently processed by Transformer layers to achieve higher-level feature interaction and integration. To further examine whether the proposed knowledge-guided mechanism provides advantages beyond simple category feature encoding schemes, a OneHot-FT-Transformer baseline model was developed. In this model, varietal category information is encoded as one-hot vectors and directly concatenated with spectral features as model inputs, serving as a reference to evaluate the effectiveness of structured knowledge integration in the proposed framework.

All deep learning baseline models were trained and evaluated using the same data partitioning strategy, loss function, and evaluation metrics to ensure fairness and consistency in performance comparison.

2.5. Quality trait fingerprint

HSI integrates spatial and spectral domains, providing distinct advantages for revealing tissue heterogeneity and visualizing the spatial distribution of plant quality traits (Wang et al., 2022). In this study, the optimized prediction models were applied pixel-wise within the defined ROI to construct quality maps. For each pixel, spectral features were extracted and input into regression models to estimate SPAD, SSC, and MC. The predicted values were subsequently transformed into pseudo-color images using a unified color scale to ensure comparability across samples. These spatial prediction maps were further organized as quality trait fingerprints, capturing intra-plant heterogeneity and revealing distinctive spatial distribution patterns among cultivars. Such fingerprints offer an intuitive interpretation of spatial quality variation and provide auxiliary references for variety identification, quality classification, and grading. Furthermore, this fingerprint-based representation establishes a scalable basis for extending hyperspectral prediction from single-value estimation to multidimensional spatial characterization.

2.6. Evaluation metrics

Model performance was primarily evaluated using the coefficient of determination (R²), the root mean square error (RMSE) and the residual predictive deviation (RPD) These metrics are defined as:

$$R^2=1-\frac{{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}$$ (8)

$$\mathit{RMSE}=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$ (9)

$$\mathit{RPD}=\frac{\mathit{SD}}{\mathit{RMSE}}$$ (10)

where n is the number of samples, $y_i$ is the observed value, ${\widehat{y}}_i$ is the predicted value, $\overline{y}$ is the mean of the observed values, and SD is standard deviation. An RPD value greater than 2.5 generally indicates that the model is suitable for NIR prediction, while an RPD above 3.0 suggests strong predictive performance.

2.7. Experimental setup

All experiments were conducted on a workstation running Windows 11 (64-bit) operating system. The hardware configuration included an Intel(R) Core(TM) i7–12,700 CPU @ 2.10 GHz, 32 GB of RAM, a 1 TB solid-state drive, and an NVIDIA GeForce RTX 3070 Ti GPU with 8 GB of dedicated memory. The software environment was built on Anaconda 3.5.0 (Anaconda Inc., USA), with CUDA 11.1 (NVIDIA, USA) and cuDNN 8.0.4 (NVIDIA, USA) for GPU acceleration. The open-source deep learning framework PyTorch served as the primary development platform, and Python 3.8.5 (Python Software Foundation, USA) was used as the programming language. Model training and code development were carried out using PyCharm 2022.01 (JetBrains, Czech Republic).

3. Results

3.1. Quality trait and spectral analysis of lettuce

Fig. 4 illustrates the average reflectance spectra of lettuce samples extracted from hyperspectral images within the 366.36–976.11 nm wavelength range, together with the results obtained after different preprocessing methods. The reflectance curves describe the optical response of plant tissues at varying wavelengths, thereby reflecting intrinsic biophysical, physiological, and biochemical properties. Overall, the spectra exhibit a typical plant reflectance pattern across the visible to near-infrared region, which has clear biological interpretations. Around 530–570 nm, a distinct reflectance variation is observed, corresponding to absorption peaks of lutein and other carotenoids (Van Wittenberghe et al., 2019). This region is closely linked to antioxidant capacity and nutritional status (Pacumbaba & Beyl, 2011). Near 700 nm, a sharp increase in reflectance forms the characteristic “red edge,” which is strongly associated with chlorophyll content and photosynthetic activity, serving as an indicator of photosynthetic health (Yu et al., 2022). In the near-infrared region, reflectance rises markedly before forming a trough around 960 nm, which coincides with the O—H bond absorption feature of water and is thus an important marker of tissue moisture content (Zhang et al., 2018). These spectral signatures are closely associated with key lettuce quality traits such as SSC, SPAD, and MC.

Fig. 4.

Average spectral reflectance of lettuce samples before and after spectral preprocessing. (a) Raw spectral reflectance. (b) Spectral reflectance after SG smoothing. (c) Spectral reflectance after MA smoothing. (d) Spectral reflectance after D1 processing.

Fig. 5 presents the distribution of quality traits across different lettuce subpopulations and their corresponding spectral reflectance characteristics. Significant inter-subpopulation variations were observed, highlighting the physiological diversity of lettuce varieties.

Fig. 5.

Distribution of quality traits and corresponding spectral reflectance characteristics of lettuce samples.

SPAD values ranged from 9.9 to 44.1, with an overall average of 28.52. Subpopulation B showed the lowest average SPAD of 25.3, indicating weaker chlorophyll accumulation, whereas subpopulations C (30.5) and R (31.6) exhibited relatively higher levels. Subpopulation W had the highest average SPAD of 41.1, suggesting stronger pigment synthesis and photosynthetic potential. These differences were consistent with spectral variations near the 700 nm red edge region.

SSC ranged from 0.900 to 5.550, with an overall average of 2.243. Subpopulation C exhibited the highest average SSC of 2.758, reflecting stronger sugar accumulation ability. Subpopulations B (2.379) and S (2.358) were intermediate, whereas subpopulation W showed the lowest average SSC of 1.800, indicating lower carbohydrate content. Spectral differences around 580 nm, related to carotenoid absorption, closely matched these SSC variations.

MC ranged from 0.7636 to 0.9880, with an average of 0.9470, indicating generally high and relatively stable water content. Subpopulations B, C, and L showed higher average MC values (0.9504), demonstrated strong water retention ability, whereas subpopulation R (0.9447) was slightly lower. Subpopulations S (0.9309) and W (0.9289) showed further reductions, while subpopulation O had the lowest average MC of 0.8954. These differences corresponded well with the reflectance trough near 960 nm, confirming the sensitivity of hyperspectral signals to moisture content.

In summary, lettuce subpopulations exhibited clear differences in SPAD, SSC, and MC, and these variations were consistently reflected in the corresponding spectral signatures. These observations highlight the intrinsic genetic and physiological diversity among lettuce varieties and provide a solid spectral and biological basis for subsequent regression modeling.

3.2. Performance analysis

To systematically evaluate the predictive performance of different modeling strategies for lettuce quality traits, comparative experiments were conducted on SPAD, SSC, and MC. The results obtained using a fixed random seed (seed = 123) are summarized in Table 1.

Table 1.

Performance comparison of different models for lettuce quality trait in prediction.

Quality trait	Model	Rt2	RMSEt	RPDt	Rc2	RMSEc	RPDc	Rp2	RMSEp	RPDp
SPAD	SPXY + None + None + SVR	0.7977	2.9145	2.22	0.7521	3.2280	2.01	0.7318	3.4806	1.93
	KS + SG + None + PLSR	0.8783	2.2606	2.87	0.7936	2.9458	2.20	0.7521	3.3461	2.01
	ACNN	0.8721	2.4202	2.80	0.8401	2.3205	2.50	0.8376	2.3416	2.48
	Spectral-ViT	0.8918	2.1420	3.04	0.8027	3.0874	2.25	0.7597	3.2945	2.04
	CNN-Transformer	0.9461	1.5112	4.31	0.8936	2.1148	3.07	0.8753	2.3736	2.83
	OneHot-FT-Transformer	0.9026	2.0225	3.20	0.8560	2.4604	2.64	0.8387	2.6991	2.49
	KG-FT-Transformer	0.9716	1.0922	5.93	0.9598	1.2998	4.99	0.9534	1.4513	4.63
SSC	SPXY + WAVE + None + SVR	0.8671	0.3225	2.74	0.8282	0.2865	2.41	0.7110	0.3537	1.86
	SPXY + None + None + PLSR	0.9052	0.2713	3.25	0.8320	0.2883	2.44	0.7298	0.3419	1.92
	ACNN	0.9046	0.2722	3.24	0.8511	0.2667	2.59	0.8412	0.2088	2.51
	Spectral-ViT	0.8572	0.3330	2.65	0.8664	0.2526	2.74	0.7571	0.3242	2.03
	CNN-Transformer	0.9235	0.2437	3.62	0.8615	0.2572	2.69	0.8585	0.1971	2.66
	OneHot-FT-Transformer	0.8986	0.2807	3.14	0.8489	0.2686	2.57	0.8318	0.2202	2.44
	KG-FT-Transformer	0.9383	0.2210	4.03	0.9352	0.1862	3.93	0.9185	0.1495	3.50
MC	Random + MA + None + SVR	0.8130	0.0243	2.31	0.7669	0.0136	2.07	0.7257	0.0123	1.91
	Random + DT + Cars + PLSR	0.8557	0.0213	2.63	0.7841	0.0131	2.15	0.7403	0.0120	1.96
	ACNN	0.8737	0.0197	2.81	0.8303	0.0116	2.43	0.8194	0.0100	2.35
	Spectral-ViT	0.8719	0.0199	2.79	0.7034	0.0145	1.84	0.7925	0.0132	2.20
	CNN-Transformer	0.9022	0.0174	3.20	0.8758	0.0099	2.84	0.8573	0.0089	2.65
	OneHot-FT-Transformer	0.8764	0.0195	2.84	0.8228	0.0118	2.38	0.8269	0.0120	2.40
	KG-FT-Transformer	0.9414	0.0075	4.13	0.9361	0.0071	3.96	0.9226	0.0066	3.60

For SPAD prediction, the best-performing traditional models were SPXY + None + None + SVR and KS + SG + None + PLSR, achieving prediction-set coefficients ${\mathrm{R}}_{\mathrm{p}}^2$ of 0.7318 and 0.7521, respectively, with corresponding RPD values of 1.93 and 2.01. These results indicate moderate predictive capability but limited predictive consistency on the independent prediction set. For SSC, SPXY + None + None + PLSR yielded the highest traditional-model performance, with ${\mathrm{R}}_{\mathrm{p}}^2$ of 0.7298 and an RPD of 1.92, while SPXY + WAVE + None + SVR showed comparable but slightly weaker generalization. For MC prediction, Random + MA + None + SVR and Random + DT + Cars + PLSR achieved ${\mathrm{R}}_{\mathrm{p}}^2$ values of 0.7257 and 0.7403, respectively, with RPD values below 2.0. Overall, although traditional methods were able to capture basic spectral-trait relationships, their prediction accuracy and generalization performance remained limited, particularly on independent prediction set.

In contrast, deep learning models exhibited substantially improved performance across all three quality traits. The ACNN model consistently outperformed traditional methods, achieving prediction set ${\mathrm{R}}_{\mathrm{p}}^2$ values of 0.8376, 0.8412, and 0.8194 for SPAD, SSC, and MC, respectively. These results demonstrate the advantage of deep learning in automatically extracting nonlinear spectral features without manual preprocessing or feature selection. The Spectral-ViT model showed competitive performance on the training and calibration sets but exhibited relatively weaker generalization on the prediction set, particularly for SPAD and MC, indicating sensitivity to sample size and a potentially increased risk of overfitting when modeling spectral data. The CNN-Transformer hybrid model further improved prediction performance by combining convolutional local feature extraction with Transformer-based global dependency modeling. Compared with ACNN, CNN-Transformer achieved higher prediction set ${\mathrm{R}}_{\mathrm{p}}^2$ values of 0.8753 (SPAD), 0.8585 (SSC), and 0.8573 (MC), indicating enhanced feature interaction modeling capability. The OneHot-FT-Transformer, which directly concatenates one-hot encoded varietal category information with spectral features, provided moderate performance improvements over ACNN but did not consistently outperform the CNN-Transformer. This result suggests that simple category feature concatenation is insufficient to fully exploit varietal information for hyperspectral regression in a data-driven manner. By contrast, the proposed KG-FT-Transformer consistently achieved the best performance across all quality traits and all dataset partitions. Specifically, prediction set ${\mathrm{R}}_{\mathrm{p}}^2$ values reached 0.9534 for SPAD, 0.9185 for SSC, and 0.9226 for MC, with corresponding RPD values of 4.63, 3.50, and 3.60, respectively. Compared with ACNN, the KG-FT-Transformer increased the prediction set ${\mathrm{R}}_{\mathrm{p}}^2$ by 0.1158, 0.0773, and 0.1032 for SPAD, SSC, and MC, respectively.

These results demonstrate that explicitly embedding varietal category knowledge into the Transformer-based feature interaction process provides clear advantages over both conventional deep learning models and naive one-hot encoding strategies. The KG-FT-Transformer effectively captures both intra- and inter-varietal spectral relationships, thereby achieving consistently higher prediction accuracy across different quality traits.

To further evaluate the statistical significance of the predictive performance differences between the proposed KG-FT-Transformer and representative deep learning baseline models, paired t-tests were conducted based on the prediction-set results obtained under multiple random seeds (see Tables S1-S3). The statistical test results are summarized in Table 2.

Table 2.

Prediction performance and paired t-test results of different models.

Quality trait	Model	Rp2	95% [CI]	RMSEp	95% [CI]	t-value (Rp2)	p-value (Rp2)	t-value (RMSEp)	p-value (RMSEp)
SPAD	ACNN	0.8477 ± 0.0225	[0.8198,0.8755]	2.2622 ± 0.1760	[2.0437,2.4807]
	CNN-Transformer	0.8679 ± 0.0092	[0.8564,0.8793]	2.4420 ± 0.0848	[2.3367,2.5473]
	KG-FT-Transformer	0.9512 ± 0.0027	[0.9478,0.9545]	1.4849 ± 0.0407	[1.4343,1.5355]
	t1					10.9540	0.000395	−10.9609	0.000394
	t2					17.5809	0.000061	−19.8258	0.000038
SSC	ACNN	0.8236 ± 0.0109	[0.8101,0.8371]	0.2200 ± 0.0069	[0.2114,0.2285]
	CNN-Transformer	0.8564 ± 0.0061	[0.8488,0.8640]	0.1985 ± 0.0041	[0.1934,0.2036]
	KG-FT-Transformer	0.9111 ± 0.0073	[0.9021,0.9201]	0.1561 ± 0.0063	[0.1483,0.1640]
	t1					16.4108	0.000081	−17.2285	0.000067
	t2					14.7824	0.00012	−14.3199	0.000138
MC	ACNN	0.8315 ± 0.0105	[0.8185,0.8446]	0.0097 ± 0.0003	[0.0093,0.0100]
	CNN-Transformer	0.8356 ± 0.0234	[0.8066,0.8646]	0.0095 ± 0.0007	[0.0087,0.0104]
	KG-FT-Transformer	0.9211 ± 0.0011	[0.9197,0.9225]	0.0066 ± 0.0000	[0.0066,0.0067]
	t1					18.9329	0.000046	−25.3247	0.000014
	t2					8.5768	0.001015	−10.4286	0.000478

t1 and t2 represent the paired t-test results between ACNN and KG-FT-Transformer, and between CNN-Transformer and KG-FT-Transformer, respectively. Statistical significance is determined based on the corresponding p-values (p < 0.01).

For the SPAD prediction task, the KG-FT-Transformer achieved an average prediction set performance of ${\mathrm{R}}_{\mathrm{p}}^2$ = 0.9512 ± 0.0027, which was significantly higher than that of both the ACNN and CNN-Transformer models. The corresponding paired t-test results were t₁(4) = 10.9540, p < 0.01 and t₂(4) = 17.5809, p < 0.01 respectively. In addition, the 95% confidence interval (CI) for ${\mathrm{R}}_{\mathrm{p}}^2$ was [0.9478, 0.9545]. Meanwhile, the prediction set RMSE of the KG-FT-Transformer was 1.4849 ± 0.0407, which was significantly lower than those of the two baseline models, with paired t-test results of t₁(4) = −10.9609, p < 0.01 and t₂(4) = −19.8258, p < 0.01 respectively, and a 95% CI of [1.4343,1.5355]. These results demonstrate the effectiveness of the proposed method for SPAD prediction. Similarly, consistent results were observed for both SSC and MC prediction tasks. In all cases, the KG-FT-Transformer achieved significantly higher ${\mathrm{R}}_{\mathrm{p}}^2$ values and significantly lower RMSE values than the ACNN and CNN-Transformer models, with all paired t-test results reaching a statistically significant level (p < 0.01).

Overall, the paired t-test results indicate that, under multiple random seed conditions, the KG-FT-Transformer consistently exhibits stable and statistically significant performance advantages across all three quality trait prediction tasks (p < 0.01). This confirms that the observed performance improvements are not attributable to random initialization effects, but rather stem from the effectiveness of the proposed model architecture.

To further verify the optimization behavior and training stability of different models beyond performance metrics, the training and calibration loss curves were visualized, as shown in Fig. S1. Across all three quality traits, the proposed KG-FT-Transformer exhibits a more gradual loss reduction during the early training stage compared with ACNN and CNN-Transformer. This slower initial decrease reflects the increased optimization complexity introduced by the integration of varietal category knowledge and structured feature interaction, where the model first aligns spectral representations with a category-aware latent space rather than aggressively minimizing empirical loss. As training proceeds, the KG-FT-Transformer converges steadily and reaches a stable loss level within a relatively small number of epochs, indicating efficient convergence and stable optimization dynamics. In contrast, although ACNN and CNN-Transformer show a rapid loss reduction at the beginning, their training typically requires a substantially larger number of epochs to stabilize, and in some cases, the calibration loss exhibits noticeable fluctuations or late-stage increases, suggesting a potential risk of overfitting. These observations indicate that, for small-sample hyperspectral regression tasks, conventional deep learning models may tend to overfit when trained for extended epochs, whereas the proposed KG-FT-Transformer achieves a better balance between convergence speed, training efficiency, and generalization stability by incorporating explicit varietal knowledge into the learning process.

Beyond training behavior analysis, the agreement between predicted and observed values was further evaluated using density scatter plots, as shown in Fig. 6. The calibration and prediction sets revealed that most points clustered tightly along the diagonal y = x, indicating strong linear agreement between predicted and observed values. For SPAD, the calibration and prediction sets achieved R² values of 0.9598 and 0.9534, respectively, while the prediction set RMSE and RPD were 1.4513 and 4.63. For SSC, the calibration and prediction sets achieved R² values of 0.9352 and 0.9185, respectively, while the prediction set RMSE and RPD were 0.1495 and 3.50. For MC, the calibration and prediction sets achieved R² values of 0.9361 and 0.9226, respectively, while the prediction set RMSE and RPD were 0.0066 and 3.60. These results indicate low prediction errors and stable performance across both calibration and prediction sets.

Fig. 6.

Prediced versus observed values and residual analysis for lettuce quality traits.

Residual analysis further validated model reliability. Most residuals were randomly distributed around the zero-error line without observable trends (e.g., curvature or funnel effects), suggesting that prediction errors primarily arose from random noise rather than systematic over- or underestimation. The vertical dispersion of residuals remained relatively constant across the prediction range, and while a few outliers existed, they did not materially affect overall performance.

Collectively, these findings indicate that the proposed KG-FT-Transformer consistently outperformed both traditional methods and the selected deep learning baseline models across all metrics. Its consistently high prediction accuracy and unbiased estimation highlight its strong potential for high-precision, non-destructive prediction of lettuce quality traits.

3.3. Ablation study

To assess the contribution of individual components in the proposed KG-FT-Transformer, a set of ablation experiments was conducted by constructing three variant models: (1) W/O GFFN, in which the GFFN was replaced with a FNN. (2) W/O KG, where the knowledge-guided (KG) module was removed to evaluate the impact of varietal category information. (3) W/O KG-GFFN, in which both the KG and GFFN modules were removed to examine their combined effect.

Fig. 7 summarizes the prediction performance of these variants across the three lettuce quality traits. Compared with the complete KG-FT-Transformer, all ablated models exhibited varying degrees of performance degradation, indicating that each component contributes to the overall effectiveness of the proposed model.

Fig. 7.

Ablation study results of different KG-FT-Transformer variants for quality trait prediction.

For the W/O GFFN variant, replacing the GFFN with a standard FNN resulted in a modest decline in prediction performance. On the SPAD prediction set, the R², RMSE, and RPD were 0.9509, 1.4897, and 4.51, respectively. For SSC, the corresponding values were 0.9153, 0.1525, and 3.44, while for MC they were 0.9218, 0.0079, and 3.58. Although performance remained relatively high, the consistent reduction compared with the full model suggests that the GFFN contributes positively to nonlinear feature transformation and efficient representation learning.

In contrast, the W/O KG variant showed a more pronounced performance decline. The prediction set R² values decreased to 0.8891 for SPAD, 0.8600 for SSC, and 0.8591 for MC, accompanied by noticeable increases in RMSE and reductions in RPD. On average, R² decreased by more than 6% relative to the complete model. These results indicate that incorporating varietal category knowledge plays a critical role in stabilizing spectral-trait modeling under limited sample conditions and in capturing structured relationships associated with varietal differences.

The W/O KG-GFFN variant exhibited the most substantial performance degradation. When both modules were removed, the prediction set R² values for SPAD, SSC, and MC decreased by 7.34%, 6.99%, and 7.28%, respectively, compared with the full model. This consistent decline across all traits highlights the complementary roles of the KG and GFFN modules in enhancing feature interaction and predictive accuracy.

Overall, the ablation results demonstrate that the knowledge-guided module serves as a key component for incorporating structured varietal information, while the GFFN further improves nonlinear modeling capacity. Their combined integration enables the KG-FT-Transformer to achieve more stable and accurate prediction of lettuce quality traits.

3.4. Visualization of quality trait fingerprint

To further uncover spatial variations in quality traits among lettuce varieties and subgroups, spectral data from each pixel within the ROI were processed by the optimized prediction models to generate pixel-wise quality trait estimations. These values were then visualized as spatial distribution maps using a standardized Jet Pseudocolor scheme, as shown in Fig. 8. This pixel-level prediction and visualization pipeline provides an intuitive means of examining spatial heterogeneity and aggregation patterns of quality traits, rather than relying solely on sample-level average values.

Fig. 8.

Spatial visualization of quality trait fingerprints in lettuce based on pixel-wise hyperspectral prediction. (a) B222-1. (b) C545-1. (c) L811-3. (d) O901-1. (e) R1106-2. (f) S1201-1. (g) W1277-3.

The visualization results indicate that high-value regions of SPAD and SSC are primarily concentrated in the central areas of the leaves. For example, variety W1277 exhibited the highest SPAD content, with its fingerprint map showing a typical central aggregation pattern. Similarly, variety C545 displayed the highest SSC content, also characterized by central enrichment. Such visually similar spatial distribution patterns suggest that, under the experimental conditions of this study, SPAD and SSC may share consistent spatial distribution characteristics at the leaf level.

In contrast, for MC, variety O901 exhibited overall pixel-wise prediction values higher than the average level observed among other varieties, with elevated MC values distributed across multiple leaf regions. This spatial pattern visually highlights a distinct water-related distribution characteristic, suggesting potential varietal differences in moisture retention behavior that can be captured by the proposed model at the pixel level.

Overall, the quality fingerprints constructed through hyperspectral pixel-wise regression and pseudocolor mapping reveal both macro-level differences in predicted trait magnitudes and micro-level spatial distribution characteristics from a qualitative perspective. The visually consistent spatial distribution patterns observed for SPAD and SSC, together with the distinct spatial patterns identified for MC, provide qualitative, spatially informed references that can assist the interpretation of DUS-related trait attributes. The proposed quality fingerprint further offers a spatial analysis basis for subsequent quantitative trait characterization, variety differentiation, and fine-scale quality assessment in lettuce.

4. Discussion

4.1. ML vs. DL under small-sample conditions

Traditional ML methods, such as PLSR and SVR, have been widely applied in hyperspectral crop quality trait prediction due to their relatively strong interpretability and stable modeling performance under limited sample sizes (Wei et al., 2025). These methods typically rely on explicit spectral preprocessing and feature selection to reduce dimensionality and mitigate multicollinearity, and are effective for modeling linear or weakly nonlinear spectral-trait relationships (Feng et al., 2024; Saha & Manickavasagan, 2021). However, when faced with high-dimensional hyperspectral data with complex structures, their modeling capacity becomes limited, and prediction performance on independent prediction sets often remains suboptimal (Song et al., 2025).

In contrast, DL models enable end-to-end feature learning directly from raw spectral inputs, eliminating the need for manual feature engineering. By automatically capturing complex nonlinear relationships and higher-order feature interactions, DL approaches show clear advantages in extracting discriminative spectral representations. Nevertheless, under small-sample conditions, conventional DL models are more susceptible to overfitting, unstable optimization, and performance degradation on unseen data. This challenge is particularly evident in hyperspectral regression tasks, where model parameter scales are large relative to the available effective samples.

Therefore, although DL models offer superior theoretical modeling capacity, their direct application to small-sample hyperspectral prediction tasks remains nontrivial. Simply increasing model complexity does not necessarily lead to better performance. Instead, effective model design must balance expressive power with appropriate structural constraints to improve sample efficiency, optimization stability, and prediction consistency.

4.2. Comparative analysis of DL models

To overcome the limitations of purely data-driven DL models under small-sample conditions, this study proposes the KG-FT-Transformer, which explicitly incorporates varietal category knowledge into the spectral feature interaction process. By embedding structured prior information into a Transformer-based representation learning framework, the model jointly captures spectral reflectance patterns and both intra- and inter-varietal feature relationships. This design enhances optimization stability and prediction consistency while maintaining high prediction accuracy. Accordingly, the proposed framework is particularly suitable for quality trait prediction tasks conducted under identical or similar experimental conditions, and provides an effective approach for rapid quality assessment and decision support in applications such as intelligent breeding, food processing, and quality control. Its predictive performance and stable training behavior further demonstrate its practical applicability.

By comparison, although ACNN and CNN-Transformer exhibit strong fitting capability during training, their prediction-set performance is relatively more sensitive to sample size and data partitioning, which can be attributed to their architectural characteristics. ACNN integrates convolutional operations with attention mechanisms to emphasize informative spectral regions and extract local spectral features. This structure enables effective modeling of nonlinear spectral-trait relationships; however, under conditions of limited sample size and increased sample heterogeneity, the learned representations tend to adapt closely to the training data distribution, thereby constraining prediction consistency. The CNN-Transformer model combines the local feature extraction ability of convolutional networks with the global dependency modeling of Transformer encoders, allowing it to capture long-range correlations among spectral features. While this hybrid architecture offers enhanced representational capacity, the inherent complexity of the Transformer component increases the risk of overfitting in small-sample scenarios. This effect is often reflected by calibration performance fluctuations or degradation during later training stages. As a result, CNN-Transformer models are more likely to achieve their full potential when larger training datasets or stronger regularization strategies are available.

Overall, the KG-FT-Transformer is better suited for hyperspectral quality prediction tasks that demand high prediction accuracy and stable performance under limited sample conditions. ACNN remains advantageous in scenarios dominated by strong local spectral patterns and relatively larger datasets, whereas CNN-Transformer models show greater potential when sufficient training data are available. These findings indicate that, for small-sample hyperspectral regression problems, the rational integration of domain knowledge and the introduction of structured constraints into DL frameworks are critical for improving robustness and practical applicability.

4.3. Quality trait fingerprints and spatially explicit interpretation

Quality trait fingerprints are constructed based on leaf-scale spatial information to provide an intuitive representation of hyperspectral prediction results. By visualizing pixel-wise trait distributions through pseudo-color mapping, these fingerprints reveal spatial expression patterns associated with trait levels, spatial uniformity, and inter-cultivar variation. Such spatially explicit representations of predicted quality traits offer complementary references for the analysis of DUS-related characteristics. By depicting the spatial distribution and heterogeneity of traits at the leaf level, the fingerprint results facilitate exploratory comparisons among cultivars under controlled experimental conditions and support qualitative interpretation of DUS-related features. The integration of hyperspectral prediction outcomes with spatial expression patterns provides a foundation for subsequent studies aimed at developing spatial descriptors and DUS-oriented indicators.

4.4. Limitations and future prospects

Although the proposed approach demonstrates promising potential, several limitations remain. First, the experiments were conducted under a single controlled environment with a relatively limited sample size, and its applicability under more diverse growing conditions has not yet been fully validated. Second, the fingerprint analysis developed in this study is primarily qualitative in nature, and its application in DUS-related evaluation requires further quantitative support.

Future research will extend the framework to larger populations and multiple growing environments to systematically assess its stability and applicability under more diverse conditions. In addition, quantitative spatial evaluation methods based on quality trait fingerprints will be explored, with the aim of strengthening the linkage between spatial prediction patterns and trait analysis or cultivar evaluation. While preserving interpretability and the non-destructive nature of the approach, these efforts are expected to further enhance the robustness and practical value of the proposed framework.

5. Conclusion

This study proposes a knowledge-guided and data-driven approach for hyperspectral prediction of lettuce quality traits, termed KG-FT-Transformer. By jointly modeling hyperspectral reflectance features and varietal category knowledge within a unified representation learning paradigm, the proposed model effectively captures spectral patterns as well as intra- and inter-varietal relationships, enabling accurate prediction of SPAD, SSC, and MC. Under controlled experimental conditions, KG-FT-Transformer consistently outperformed the selected baseline models, achieving R² values of 0.9534, 0.9185, and 0.9226, with corresponding RPD values of 4.63, 3.50, and 3.60, respectively. These results demonstrate its effectiveness for non-destructive crop quality prediction. In addition, quality trait fingerprints deriverd from pixel-wise hyperspectral predictions provide a spatially explicit representations of trait expression, offering qualitative references for DUS-related trait interpretation and auxiliary support for variety evaluation and germplasm identification.

Future work will focus on extending this approach to larger populations and more diverse growing environments. Its applicability will also be explored in phenotypic analysis of major staple crops, such as maize, wheat, and soybean. These efforts are expected to further enhance its practical relevance in agricultural science and technology.

CRediT authorship contribution statement

Guangjie Qiu: Writing – review & editing, Writing – original draft, Visualization, Methodology, Conceptualization. Xiaoqian Chen: Resources, Formal analysis, Data curation. Anran Song: Writing – review & editing, Methodology. Si Yang: Resources, Methodology. Xinyu Guo: Conceptualization, Supervision, Funding acquisition. Chunjiang Zhao: Validation, Project administration, Funding acquisition.

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

Supplementary material 1

Data availability

Data will be made available on request.