This study employs five independent datasets ( Table 1 ) from distinct ecosystems, each containing measurements of CHL, EWT, LMA, and corresponding leaf reflectance spectra. All spectral data were acquired using a high-resolution spectroradiometer (1 nm resolution). The LMA, CHL, and EWT values in each dataset were determined following standardized protocols, with LMA calculated as the ratio of leaf dry mass to projected area. Dataset 1 (D1) comprises data from five forest plant species predominantly distributed in temperate regions. Dataset 2 (D2) contains six tropical plant species spanning subtropical and tropical ecosystems. Dataset 3 (D3) includes two herbaceous species cultivated under controlled laboratory conditions. Dataset 4 (D4) incorporates two xerophytic plant species adapted to arid environments with water-limited growth conditions. These four datasets (D1–D4) are publicly available and can be accessed through the EcoSIS (Ecological Spectral Information System) platform at https://ecosis.org.

Dataset 5 (D5) was collected from an agricultural research station in Xinxiang, Henan Province, China (113°45′40″E, 35°8′11″N) ( Figure 1a ), and comprises three crop species: potato, soybean, and maize. Among these, soybean and maize were part of breeding trials and were planted in separate plots, with each plot corresponding to a unique cultivar, including 36 cultivars for maize and 151 for soybean (Figure 1b). Notably, maize was sown in two separate batches, 28 days apart, resulting in distinct growth stages between the two groups of maize plots (Figure 1d). All crops were managed according to local agricultural practices, with optimal fertilization, pest control, and field maintenance applied. Spectral measurements were conducted on fresh leaves using standardized protocols to ensure data quality and comparability.

Spectral measurements were conducted using an ASD FieldSpec3 spectroradiometer (Analytical Spectral Devices Inc., Boulder, CO, USA) (Figure 1c), which operates over a spectral range of 350–2500 nm with a 1 nm sampling interval. Prior to data collection, the instrument was calibrated with a Spectralon white reference panel to ensure accurate reflectance measurements. All measurements were performed between 10:00 AM and 2:00 PM local time under clear-sky conditions to minimize the effects of solar angles and atmospheric variability. For each crop species, fully expanded, healthy leaves were selected from the upper canopy to maintain consistency and physiological relevance. The adaxial surface of each leaf was positioned perpendicular to the optical fiber probe, and a leaf clip equipped with an internal light source was used to provide stable illumination and reduce ambient light interference. For each leaf, three replicate spectral measurements were acquired and averaged to reduce noise and enhance signal reliability.

CHL was determined using a solvent extraction method with 95% ethanol, and pigment concentrations were quantified spectrophotometrically based on absorbance at 649 nm and 665 nm, following established protocols. Leaf fresh weight and area were measured using an electronic balance and a leaf area meter, respectively. Samples were then oven-dried at 65°C for 48 hours to determine dry mass. LMA was calculated as the ratio of dry mass to leaf area, while EWT was computed as the difference between fresh and dry mass, normalized by leaf area. A total of 490 leaf samples were collected for analysis, and the proportion of samples across different crops is shown in Figure 1b .

The distribution patterns of CHL, EWT, and LMA are illustrated in Figure 2 . CHL exhibited significant variation among the datasets, with the highest mean value observed in D5 and the lowest in D4. A similar pattern was observed for EWT, where D5 showed the highest values, while D1 and D4 had comparatively lower values. In contrast, LMA exhibited a distinct distribution trend, with D4 showing the highest values and D3 the lowest, indicating substantial variability in LMA across datasets. Additionally, the correlation coefficients among the traits differ: EWT and LMA exhibit a strong positive correlation (0.66), CHL and LMA demonstrate a moderate negative correlation (-0.31), and CHL and EWT show a relatively weak negative correlation (-0.15).

This study proposes a PROSPECT Pre-training Adversarial Domain Adaptation Network (PPADA-Net), which employs a two-stage training strategy for plant trait prediction ( Figure 3 ). The core idea is to fully leverage large-scale simulated spectral data generated by the PROSPECT physical model for pretraining, followed by adaptive fine-tuning across domains using a limited number of target-domain samples.

The PROSPECT radiative transfer model (version PROSPECT-D) has been employed to generate high-fidelity leaf spectral simulation data (Feret et al., 2017). Based on physical optics principles, the PROSPECT model simulates leaf reflectance and transmittance spectra across 400–2400 nm by coupling anatomical structure parameters with biochemical parameters. Through radiative transfer equations in layered media, this model quantitatively characterizes multiple scattering and absorption effects within leaves, effectively capturing the nonlinear influence of various plant traits on spectral responses. A full-spectrum simulation dataset was established using the PROSPECT-D model to generate 20,000 synthetic spectral-trait pairs for pre-training. The model simulates leaf reflectance and transmittance across 400–2400 nm at 1 nm resolution, based on biophysical parameters: CHL (0.1–100 µg/cm²), leaf water depth (0.01–0.05 cm), and dry matter content (0.004–0.009 g/cm²), with additional modulators like total anthocyanin content (1.2-1.8) and leaf structure index (1.0–1.9). Leaf water depth and dry matter content can be converted to EWT and LMA, respectively, enabling direct mapping between physical parameters and functional traits. A Latin hypercube sampling strategy ensured uniform parameter sampling, covering biologically plausible ranges ( Table 2 ). Each sample comprises a reflectance spectrum (400–2400 nm) paired with corresponding CHL, EWT, and LMA values, mimicking diverse ecosystem conditions.

The simulated spectra are fed into a ResNet-based encoder network to extract high-level spectral features ( Figure 4 ). The architecture processes spectral inputs through a hierarchical feature extraction pipeline. The hyperspectral reflectance data consisted of 2000 bands (400–2400 nm at 1 nm resolution). To enable compatibility with ResNet, the 1D spectra were reshaped into 2D arrays of size 224×224 and duplicated across three channels to form 224×224×3 tensors. This operation preserves spectral information and allows the use of 2D convolutional layers, which are effective in capturing local and hierarchical patterns. This transformation is purely structural and preserves the original spectral information. The network commences with a 7×7 convolutional layer (64 filters, stride=2) followed by 3×3 max-pooling, establishing preliminary spatial-spectral representations. Subsequent residual blocks employ bottleneck structures with cascaded 1×1 and 3×3 convolutions, progressively expanding channel dimensions from 64 to 1024 through four major stages. Each stage contains multiple identity-short cut blocks where 1×1 convolutions perform channel dimension matching, while 3×3 convolutions extract spatially invariant features. Notably, the architecture implements channel scaling factors of ×4 between stages (64→256→512→1024), maintaining computational efficiency through bottleneck compression. The deep stack of 21 convolutional layers leverages residual connections to preserve gradient flow, with feature map spatial resolution systematically reduced through stride convolutions in transitional blocks. This design enables effective learning of multiscale spectral-spatial correlations while mitigating vanishing gradient issues inherent to deep networks. At this stage, a fully connected regression head is appended to the encoder output. The final encoder outputs high-dimensional latent representations suitable for downstream regression tasks through attached task-specific heads. We optimize the network by minimizing the mean squared error (MSE) loss function (LeCun et al., 2015) to predict LMA, EWT, and CHL values, with gradient backpropagation throughout the network to learn latent representations strongly correlated with plant traits from the large-scale simulated data. The MSE LOSS function is defined in Equation 1.

where m represents the trait index, and M represents the number of samples. The loss for combining all traits is defined in Equation 2.

Following pre-training, the second stage involves jointly feeding both the source domain (real-world dataset for training) and limited target domain (real-world dataset for prediction) into the pre-trained encoder. This strategy leverages the spectral feature representation capabilities acquired from large-scale simulated data while enabling fine-tuning and domain alignment through limited target domain samples, thereby enhancing prediction accuracy in real-world target environments. To simultaneously address plant trait prediction and domain adaptation, we introduce two parallel task heads at the output of the pre-trained encoder: (1) A Multilayer Perceptron (MLP) for regressing three plant traits (LMA, EWT, and CHL) as the downstream task, optimized using MSE loss. (2) A Hierarchical Cross-Domain Feature Alignment (HCDFA) head, this module takes input feature maps from both the source domain and target domain and processes them through a two-step mechanism. First, a Contrastive Attention Module calculates the cosine similarity between the input feature maps, emphasizing shared patterns while suppressing domain-specific noise. This process generates weighted feature maps for both domains. Subsequently, an adversarial contrastive loss ACL is applied to refine these weighted feature maps, increasing the similarity between the aligned feature representations of the source and target domains. These similarity weights are applied to the source domain feature map, dynamically adjusting its feature distribution to produce the aligned source domain feature map. Simultaneously, the target domain features undergo adaptive mapping to generate the aligned target domain feature map. During forward propagation, the HCDFA acts as an identity transform, while in backward propagation it inverts gradients from the DA loss, thereby encouraging the encoder to learn domain-invariant features through adversarial confusion. The composite objective function during this stage comprises two components is defined in Equation 3.

L M S E represents the MSE loss for plant trait regression, while L D A denotes the domain adaptation loss. λ₁ and λ₂ are hyperparameters that control the weighting of the two loss terms, and their optimal balance can be determined through hyperparameter tuning on the validation set.

During network training, stochastic gradient descent is used to update the parameters of both the encoder and the two task heads in mini batches. The Few-shot learning strategy incorporates a small amount of labeled target-domain data into the network alongside source-domain data. A feature alignment mechanism mitigates domain shift, facilitating cross-domain knowledge transfer between datasets. Model training was conducted using a staged optimization strategy. In the first stage, pre-training was conducted using PROSPECT-D-generated spectral-trait pairs to initialize the ResNet-based encoder. The network, structured with a 7×7 convolutional layer (64 filters, stride=2), 3×3 max-pooling, and residual blocks scaling channels from 64 to 1024, was trained for 200 epochs with a batch size of 64. The Adam optimizer was used with an initial learning rate of 0.001 for rapid convergence, decaying by 1×10–5 in later stages to prevent overfitting. The MSE loss function guided optimization, enabling the encoder to learn spectral features strongly correlated with CHL, EWT, and LMA, forming a robust foundation for subsequent domain adaptation. In the second stage of transfer learning, under the integration of source and target domain data, the number of epochs was adjusted to 100 with a batch size of 32, and the initial learning rate was updated to 1×10^-4 to facilitate more refined fine-tuning and domain adaptation.

Four regression methodologies were systematically compared for CHL, EWT, and LMA prediction: (1) conventional multivariate PLSR, (2) data-driven ResNet, along with two ablated variants, (3) physics-enhanced ResNet-PROSPECT and (4) domain-adaptive ResNet-GRL. These were rigorously benchmarked against our proposed PPADA-Net to evaluate the incremental benefits of integrated physical-adversarial learning.

As a conventional multivariate statistical approach, PLSR establishes linear relationships between hyperspectral reflectance (predictors) and plant traits (response variables) through latent variable decomposition. This model maps the relationship between spectral data and plant traits to a low-dimensional latent space, allowing PLSR to perform effective regression analysis while maintaining low computational cost, representing traditional chemometrics methodology in spectral analysis (Wold et al., 2001).

The standard ResNet18 architecture was adapted for spectral regression, employing residual blocks with 1D convolutions to capture hierarchical spectral features. Without any domain adaptation mechanisms, this deep learning baseline utilized raw spectral inputs (400–2400 nm) to directly predict trait values through fully connected regression layers, demonstrating pure data-driven modeling capability (Chen et al., 2022).

ResNet-PROSPECT integrated physics-informed pretraining by initializing weights through simulated data generated from the PROSPECT-D radiative transfer model. The network first underwent 100-epoch pretraining on 10000 synthetic spectra-trait pairs covering the full parameter space, followed by fine-tuning on experimental datasets, testing the isolated effect of physical prior integration.

The ResNet-GRL architecture implements domain adversarial learning without physical constraints by appending a GRL between the feature extractor and the domain classifier. This dual-objective network simultaneously minimizes trait prediction error and maximizes domain confusion through adversarial training (λ = 0.3), thereby evaluating the independent contribution of domain-invariant feature learning.

Two distinct validation paradigms were implemented to comprehensively assess model performance: (1) Aggregated Cross-Validation ( Figure 5a ): All samples from the five datasets (D1-D5) were pooled into a composite repository, followed by a stratified five-fold cross-validation scheme, where each fold maintained proportional representation of the original dataset distributions. In each iteration, 80% of the samples were allocated for training (with 15% reserved for internal validation to facilitate early stopping) and 20% for testing. This approach assessed general predictive accuracy under the assumption of homogeneous data distribution. (2) Cross-Dataset Validation ( Figure 5b ): To rigorously evaluate cross-domain transferability, leave-one-dataset-out experiments were conducted. In each trial, four datasets were used as the training set, while the remaining dataset was held out for independent testing. Spatial-spectral standardization was applied to each dataset using the respective training statistics to prevent information leakage. This protocol specifically quantified the model’s generalization capacity across heterogeneous data acquisition conditions. (3) Training on D1–D4 with Testing on D5 ( Figure 5c ): Datasets D1–D4 were used as the training set, and the model’s predictive performance was tested on the independent, independently collected field dataset D5. This validation methodology is particularly valuable for assessing model robustness and practical applicability in real-world agricultural settings.

In the model evaluation, to measure the variability of the dependent variable and prediction error, the normalized root means square error (nRMSE) and the coefficient of determination (R²) are formulated in Equations 4, 5.

where n is the number of samples; y i and y ^ i represent the measured and the predicted grain yield of sample i, respectively. A higher value of R² and lower nRMSE values would indicate superior performance of the model.

Statistical analyses of spectral reflectance across the 400–2400 nm range are illustrated in Figure 6a . Consistent patterns were observed across the five datasets, with all exhibiting identifiable spectral peaks and troughs. In the visible region (500–700 nm), mean reflectance values remained low and relatively stable, reflecting strong pigment absorption. A sharp increase occurred in the near-infrared (NIR, 700–1300 nm), consistent with the high reflectance associated with internal leaf structure. This was followed by a fluctuating decline in the short-wave infrared (SWIR, 1300–2400 nm), where a pronounced dip was observed between 1900 and 2100 nm—likely due to strong water absorption features. Inter-dataset differences were especially evident in specific spectral bands. For instance, the standard deviation of reflectance was notably higher in the visible spectrum, indicating greater variability among datasets in this region. Conversely, variability declined in the NIR and SWIR regions, particularly around 1000–1300 nm and 1900–2100 nm, as shown by the narrower spread in standard deviation and coefficient of variation (CV) curves.

Figure 6b presents the Pearson correlation between spectral reflectance and plant traits. For CHL, the strongest correlations were found in the visible spectrum (400–700 nm), aligning with known pigment absorption features. Correlation strength diminished in the NIR and SWIR regions, generally falling below 0.4 except in the 1900–2400 nm sub-region. In contrast, EWT exhibited weak associations in the visible range, but showed moderate to strong correlations (r > 0.5) throughout the 700–1900 nm range, underscoring its sensitivity to water-related absorption features in those bands. For LMA, positive correlations peaked in the 700–1400 nm range, suggesting a strong relationship between biomass-related traits and reflectance in this region. Trait–spectrum relationships also varied across datasets. For example, Dataset D2 exhibited higher LMA correlations in the NIR, while Dataset D5 consistently showed weaker correlations for EWT, reflecting the combined effects of biological variability and environmental conditions. These results emphasize the importance of accounting for both spectral region characteristics and dataset heterogeneity in trait modeling.

The performance of the PLSR model and the proposed PPADA-Net in predicting three plant traits was evaluated ( Figure 10 ) across various training-testing dataset combinations (Tables 3, 4). PPADA-Net consistently outperformed the PLSR model across most dataset combinations for all three traits. On average, PPADA-Net exhibited higher R² values and lower nRMSE values than PLSR. For instance, in predicting CHL, PPADA-Net achieved a mean R² of 0.53 and a mean nRMSE of 0.12 across all dataset combinations, whereas PLSR attained a mean R² of 0.18 and a mean nRMSE of 0.25. Similarly, PPADA-Net achieved an average R² of 0.72 and an nRMSE of 0.10 in predicting LMA, whereas PLSR attained an R² of 0.46 and an nRMSE of 0.15. This indicates that PPADA-Net, which integrates simulated data pre-training with adversarial domain alignment, significantly enhances prediction accuracy and reliability.

The model’s performance varied considerably across different dataset combinations. For example, in predicting LMA, when using D1 as the training set and D2 as the testing set, PPADA-Net achieved an R² of 0.68 and an nRMSE of 0.11, whereas PLSR recorded an R² of 0.34 and an nRMSE of 0.15. Similarly, for the D5–D4 combination, PPADA-Net attained an R² of 0.81 and an nRMSE of 0.07, in contrast to PLSR’s R² of 0.22 and an nRMSE of 0.22. These examples further highlight that the prediction accuracy of data-driven models is significantly influenced by the training set. Under these conditions, PPADA-Net demonstrated greater stability and adaptability across various dataset combinations. For example, in predicting EWT across different dataset combinations, PLSR exhibited an R² range of 0.18–0.58 and an nRMSE range of 0.13–0.24. In contrast, PPADA-Net displayed an R² range of 0.53–0.82 and an nRMSE range of 0.07–0.12, indicating that the integration of simulated data pre-training and adversarial domain alignment enables the model to better handle domain shifts and maintain consistent performance across diverse data sources, which is particularly valuable in cases of high data heterogeneity. Overall, among the three traits, CHL is generally more challenging to predict accurately compared to EWT and LMA. Although PPADA-Net achieved relatively high R² values for LMA in certain cases, some dataset combinations still yielded low R² values, suggesting that achieving high-precision transfer for CHL prediction across different datasets remains challenging.

When training the model on D1–D4 and testing on the independently collected field D5 dataset, notable performance differences were observed across models. The PPADA model ( Figure 11a ) achieved high prediction accuracy for all three traits, with R² values of 0.72 (CHL), 0.78 (EWT), and 0.87 (LMA), and corresponding nRMSE values of 0.08, 0.08, and 0.07. The scatter plots show strong linear agreement between predicted and observed values, especially for LMA, where most points closely align with the 1:1 line, indicating superior predictive performance. CHL and EWT predictions were more concentrated in the low-to-mid value ranges, with slightly reduced agreement at higher values. In contrast, the PLSR model ( Figure 11b ) produced lower overall accuracy, with R² values of 0.53, 0.56, and 0.71 for CHL, EWT, and LMA, and nRMSE values of 0.10, 0.11, and 0.10, respectively. The scatter plots display greater dispersion, particularly for CHL and EWT, where deviations were more pronounced in the lower measurement ranges. Some improvement in agreement was observed at higher values. Among the three traits, LMA still yielded the best performance, but compared to PPADA, its R² was 0.16 lower and nRMSE was 0.03 higher.

Accurate mapping of plant functional traits serves as a pivotal tool for advancing crop breeding and precision agricultural management. The PPADA model was trained using datasets D1-D4 and subsequently applied to spatially visualize CHL, EWT, and LMA in 490 field samples from the independent test set D5, as shown in Figure 11 . The results demonstrate a high degree of consistency between the spatial maps predicted by PPADA and the ground truth measurements, validating the robustness of the model in trait estimation. Among the three traits, LMA exhibited the highest spatial consistency, with prediction errors concentrated in the low-value range. For CHL and EWT, although most predictions closely matched the observed values, the model showed slight underestimation for samples with exceptionally high values. Overall, the PPADA model demonstrated its capability for high-throughput, spectral-based trait monitoring and field mapping applications.

The use of spectral features, including reflectance data, vegetation indices, and spectral derivatives, for plant trait prediction has been widely applied in previous studies (Li et al., 2024). For example, Aguirre-Gutierrez et al. (2021) demonstrated that key plant functional traits can be accurately predicted across the tropics using the high spatial and spectral resolution of Sentinel-2 imagery in conjunction with climatic and soil information. While PLSR offers high interpretability and computational efficiency in plant traits estimation, it has several limitations. First, it does not support multi-task learning, requiring separate models for different traits, which results in fragmented workflows. Second, it struggles to capture complex spectral-trait relationships, especially in high-dimensional or nonlinear contexts, such as SWIR interactions with leaf dry matter. As shown in our results, PLSR exhibited the lowest performance (mean R² = 0.59–0.72) and showed a significant decline in accuracy in cross-dataset scenarios (e.g., CHL R² = 0.18), highlighting its vulnerability to data heterogeneity.

Deep neural networks, such as CNN, have improved traits prediction by enabling end-to-end multi-task learning and hierarchical feature extraction. Cherif et al. (2023) demonstrated that CNN-based architectures could simultaneously predict multiple traits with enhanced nonlinear modeling. While these models outperform traditional methods in single-domain settings, their heavy reliance on large, labeled datasets poses a challenge (Chen et al., 2023), as training robust models requires extensive field measurements, which are costly and scarce for certain traits. Furthermore, these models often exhibit poor cross-dataset generalizability due to domain shifts in spectral patterns (Zhang and Bao, 2022). Our results support these findings: ResNet showed inconsistent performance across traits (e.g., EWT R² = 0.61 vs. PLSR’s 0.63) and experienced significant accuracy declines in transfer learning scenarios (e.g., LMA R² = 0.68 for D1-D2), highlighting its sensitivity to domain-specific biases. To overcome these limitations, PPADA-Net integrates physical priors from radiative transfer models (e.g., PROSPECT-simulated spectra) with adversarial domain adaptation, achieving dual benefits: physics-informed feature learning and robust cross-domain generalization. It outperformed all baseline models across traits (CHL: R² = 0.72; EWT: R² = 0.77; LMA: R² = 0.86), with accuracy gains of 5.1% to 22.4% over PLSR and ResNet. These results have practical implications for remote sensing-based trait monitoring in real-world settings. PPADA-Net’s strong performance in cross-dataset validation (e.g., LMA R² = 0.72 vs. PLSR’s 0.46) demonstrates its ability to mitigate domain shifts—a key challenge when applying models across regions, time periods, or sensors. This is particularly valuable for large-scale agricultural monitoring and ecosystem management, where collecting labeled data in every new condition is impractical. Furthermore, the integration of physical priors reduces reliance on field data, potentially lowering the cost and labor of trait estimation.

Physical models, such as the PROSPECT-D radiative transfer model, simulate spectral reflectance based on the biochemical and structural properties of plant leaves, providing mechanistic insights into light-matter interactions (Peters and Noble, 2020). For example, Bhadra et al. (2024) used PROSPECT to simulate hyperspectral responses under varying parameters, demonstrating its effectiveness in controlled experimental settings. While these models offer a rigorous framework for understanding spectral-trait relationships, their standalone application is constrained by computational complexity, sensitivity to input parameter accuracy, and limited adaptability to real-world environmental variability. In contrast, purely data-driven models excel at capturing complex patterns from large datasets but often lack interpretability and struggle with generalization when data is scarce or subject to domain shifts (Hou et al., 2024). Integrating physical models with data-driven approaches has emerged as a promising strategy for improving plant trait estimation.

Integrating PROSPECT-D-generated simulations with empirical datasets mitigates these limitations by leveraging the complementary strengths of both approaches. Physical models enhance training data diversity by synthesizing spectra across a broad range of trait values and environmental conditions, reducing reliance on costly field measurements. Meanwhile, data-driven methods refine feature representations and optimize nonlinear mappings, compensating for the simplifications inherent in physical models. In this study, pretraining ResNet on PROSPECT-D-simulated spectra (ResNet-PROSPECT) significantly improved prediction accuracy compared to the baseline ResNet. For instance, LMA prediction R² increased from 0.72 (ResNet) to 0.80 (ResNet-PROSPECT), while EWT accuracy improved by 17.7% (R²=0.74 vs. 0.63), highlighting the benefits of physics-informed initialization. These improvements result from two key mechanisms: (1) simulated data introduced reflectance variations under extreme or rare trait values, enhancing model robustness; (2) PROSPECT-D’s parameterization guided the network to focus on wavelengths critical for specific traits. The success of PROSPECT-D-enhanced models suggests a scalable approach for trait estimation in heterogeneous environments. By generating synthetic spectra for underrepresented conditions, physical models can pre-train networks to generalize beyond empirical dataset limitations. Future work could explore dynamically integrating physical simulations with domain adaptation, such as iteratively updating PROSPECT-D parameters based on field observations to refine synthetic data quality. Such advancements would further bridge the gap between theoretical modeling and empirical applications, fostering robust solutions for ecosystem monitoring under climate change.

The transferability of spectral-trait prediction models is crucial for real-world applications, as ecosystems exhibit substantial variability in spectral signatures and plant trait distributions due to differences in species composition, environmental conditions, and measurement protocols (Zhang et al., 2025a). Traditional machine learning and deep learning models often struggle to generalize across such heterogeneous datasets, limiting their effectiveness in large-scale monitoring. For instance, PLSR, while computationally efficient, demonstrates poor adaptability to spectral heterogeneity. In cross-dataset validation, PLSR achieved a mean R² of only 0.18 for CHL and 0.46 for LMA, with nRMSE values reaching 0.25. Similarly, deep learning models such as ResNet, despite their capacity for multi-task learning, remain susceptible to domain-specific biases.

Adversarial domain adaptation has emerged as a powerful strategy for aligning feature distributions across domains. For example, Ma et al. (2022) proposed a domain adaptation scheme named adversarial entropy optimization to learn domain-invariant features, achieving state-of-the-art performance across diverse domain adaptation tasks. In this study, PPADA-Net integrates this technique with physical priors to address spectral heterogeneity. The model achieved a mean R² of 0.53 for CHL, 0.72 for LMA, and 0.65 for EWT across all dataset combinations, outperforming PLSR by 35–52% in R². Notably, PPADA-Net maintained stable performance even in challenging scenarios, such as the D5→D4 transfer for LMA (R²=0.81 vs. PLSR’s 0.22), demonstrating its robustness against domain-specific noise. The enhanced transferability arises from two synergistic mechanisms. First, PROSPECT-D-simulated spectra provides trait-specific spectral patterns, enabling the model to prioritize domain-invariant biochemical signals over dataset-specific artifacts. Second, by minimizing discrepancies between source and target domains in high-dimensional feature space, PPADA-Net mitigates overfitting to local spectral variations. Future research could extend this framework to dynamically adapt to emerging ecosystems or sensor types, further bridging the gap between controlled simulations and field applications.

PPADA-Net demonstrated exceptional performance on the independently collected field dataset (D5), comprising potato, soybean, and maize crops under diverse cultivars and growth stages. The model achieved R² values of 0.72 (CHL), 0.78 (EWT), and 0.87 (LMA) with nRMSE reductions of 20%–30% compared to PLSR ( Figure 10 ). Notably, LMA prediction exhibited the highest accuracy (R² = 0.87), likely due to its strong spectral-physical linkage with dry matter content, as captured by PROSPECT-D simulations. These results underscore the framework’s ability to generalize beyond controlled experimental conditions, addressing the critical challenge of domain shifts caused by cultivar diversity, growth stage variability, and field-specific environmental factors. From an agronomic perspective, the high accuracy of PPADA-Net in mapping LMA and EWT ( Figure 12 ) holds significant promise for optimizing irrigation scheduling and nutrient management. For example, spatially resolved LMA estimates could guide breeders in selecting drought-tolerant cultivars, while EWT monitoring may improve water-use efficiency in water-scarce regions. However, slight underestimation of CHL in high-value ranges ( Figure 10a ) suggests that chlorophyll’s nonlinear spectral interactions under saturating conditions require further refinement. Future work should also validate the framework across broader agro-climatic zones and crop phenological stages to ensure scalability.

While PPADA-Net demonstrates significant advancements in plant trait estimation, its performance, like other machine learning methods, remains heavily dependent on the quality and quantity of training data. The framework’s effectiveness diminishes when training data are scarce or unrepresentative of target environments. For instance, PPADA-Net’s reliance on PROSPECT-D-generated simulations introduces biases if the radiative transfer model fails to capture extreme environmental conditions. In such cases, synthetic data may inadequately represent real-world spectral-trait relationships, limiting generalization. To mitigate this, future work should incorporate anthropogenic constraints into simulations, such as expanding parameter ranges for leaf structure and biochemical composition, to enhance the diversity and representativeness of synthetic datasets. This would ensure simulated spectra encompass greater information entropy, better aligning with the variability observed in field data. On the other hand, although adversarial learning in PPADA-Net effectively aligns feature distributions across datasets, its performance may degrade when domain discrepancies are extreme. While our results demonstrated robustness in most cross-dataset validations, scenarios involving stark domain shifts require careful hyperparameter tuning to stabilize adversarial training. In summary, by refining the constructive interaction between physically based simulation and adaptive learning, PPADA-Net can evolve into a reliable tool for precision agriculture.