The leaf reflectance spectra of rice seedlings under normal growth conditions (control group) and exposed to microplastic (PET, PS, and PVC) stress at two concentrations (10 mg/L and 100 mg/L) are shown in Figure 3 . The overall spectral trends of different treatment were similar and spectral peaks overlapped, which might be attributed to the dominant influence of leaf characteristics. In the blue-violet absorption region (425–490 nm), the peak of 480 nm is associated with the absorption of blue-violet light by leaves and the promotion of chlorophyll synthesis. Within the yellow-green absorption band (490–570 nm), the reflectance was observed to first increase and then decrease, with higher reflectance recorded in the green light region near 540 nm. This phenomenon was correlated with chlorophyll content in leaves and simultaneously explains the green coloration of rice leaves. The descending region at 530–660 nm, referred to as the “green edge”, was identified as a crucial indicator of chlorophyll absorption characteristics. The rapid reflectance increase in the 660–770 nm range, termed the “red edge”, was recognized as a reflection of carotenoid-related properties (Zolotukhina et al., 2024). In the near-infrared band (770–965 nm), reflectance characteristics were primarily influenced by the internal cellular structure of plants. Physiological and biochemical differences in rice leaves caused by different microplastic stresses were captured by visible-near infrared spectroscopy, establishing a technical foundation for microplastic stress detection.

To investigate the clustering distribution of rice leaves under MPs stress, PCA was performed to analyze the spectral characteristics under gradient concentrations of PET, PS, and PVC MPs. The dimensionality reduction results and corresponding loading curves are presented in Figure 4 . The cumulative contribution rates of the first three principal components reached 96.62%, 96.11%, and 96.93% for PET, PS, and PVC respectively, accounting for the majority of the variances in the original spectral data. The substantial overlap observed under different microplastic stress levels may be attributed to the subtle spectral features induced by MPs. In the PC₁ vs. PC₃ plots of PET MPs, the clustering distribution exhibited a trend of shifting from the upper right to the lower left as the stress concentration increased. Meanwhile, in the PC₁ vs. PC₂ plots of PS MPs, a clockwise rotation pattern was showed with increasing concentrations. In contrast, a counterclockwise rotational trend was exhibited as stress levels increased in the PC₁ vs. PC₂ plots of PVC MPs.

The loading weights of the principal components reflect the relative importance of spectral features for clustering and indicate specific molecular and physiological changes in leaf tissues under microplastic stress. Under the three types of MPs stress, the loading curves exhibited generally consistent patterns. In the PC₁ loadings, a trough at 660 nm corresponds to the strong absorption of chlorophyll a in the red spectral region (Zolotukhina et al., 2024). The increased reflectance may be related to chlorophyll degradation, potentially caused by MPs disrupting chloroplast membrane stability or inducing reactive oxygen species accumulation. The peak at 690 nm may reflect enhanced light scattering due to damage to the cell wall. The feature at 747 nm corresponds to the third overtone of the O–H bond, possibly indicating disruption of aquaporin formation in lipid membranes. In the PC₂ loadings, the feature at 948 nm may arise from the third overtone of N–H bonds in proteins, suggesting that MPs stress may inhibit enzyme synthesis. In PC₃, 505 nm is associated with absorption of carotenoids (Sytar et al., 2017), reflecting pigment changes induced by MPs. The features around 876 nm and 946 nm may be related to functional group interactions, such as the ester groups in PET or aromatic rings in PS, with polysaccharide O–H groups in plant cell walls. Although PCA provides a preliminary understanding of the clustering trends and associated spectral features, further methods development is needed to quantify stress levels more precisely.

The appropriate neural network structure and parameter have a significant impact on both detection performance and computational efficiency. There, the number of hidden layers and units per layer in the SE-LSTM model were systematically optimized. Based on empirical knowledge, the learning rate was set to 0.01 and the optimizer to Adam. PET MPs spectral data were as input, grid search algorithm was to explore different hidden layer configurations. The accuracy rates and loss values of the calibration set and prediction set under different hidden layer structures are shown in Figure 5 , and the hidden layers 1, 2, and 3, as well as the number of hidden layer nodes 16, 32, 64, and 128, are discussed in detail. When the model had a single hidden layer, increasing the number of nodes led to gradual improvements in accuracy and reductions in loss, with stable trends indicating enhanced performance. For models with two hidden layers, accuracy improved with more nodes, peaking at 64 units, but further increases caused accuracy to decline and calibration - prediction set gaps to widen—signs of overfitting. A similar overfitting trend was presented in models with three hidden layers as the number of units increased. Overall, the configuration with two hidden layers and 32 units was optimal, achieving high accuracy, maintaining low loss value, and offering a good trade-off between performance and generalization.

Learning rates critically govern training stability and convergence dynamics, and loss functions fundamentally determine prediction-to-target fidelity. The compelling interdependence between these parameters necessitates dedicated research to strategically select optimizers and learning rates that maximize both convergence speed and final model accuracy. In this study, the performance of three commonly used optimizers (SGD, Adam, and RMSprop) was compared across five learning rates (0.001, 0.005, 0.01, 0.05, and 0.1), as illustrated in Figure 6 . The performance of the SGD optimizer improved steadily with increasing learning rates, achieving a prediction set accuracy of 77.23% and the lowest loss of 0.55 at learning rate of 0.1. For the Adam optimizer, model accuracy initially increased and then decreased as the learning rate rose, while the loss value showed the opposite trend—decreasing first and then increasing. The optimal performance was reached at a learning rate of 0.01, where the prediction accuracy reached a maximum of 93.88% and the loss dropped to a minimum of 0.33. RMSprop exhibited a performance pattern similar to Adam but reached its peak at a learning rate of 0.005, with a validation accuracy of 84.5% and a loss of 0.53. Overall, the Adam optimizer combined with a learning rate of 0.01 demonstrated the best balance of accuracy and loss, which were selected for subsequent model training.

To comprehensively evaluate model performance, the accuracy and overfitting degree of three models on the prediction set were compared, as shown in Figure 7 . The PLSDA model achieved accuracies of 91.43%, 93.88%, and 94.33% across the three datasets, with corresponding overfitting values of 3.30%, 2.32%, and 2.57%. Although the PLSDA model exhibited relatively high accuracy, it also holed the highest overfitting values, indicating weaker model stability. The LSTM model yielded accuracies of 86.19%, 84.27%, and 87.78%, with overfitting values of 2.45%, 2.37%, and 1.71%, respectively. While its accuracy was slightly lower than that of PLSDA, the lower overfitting values suggest better generalization capability. The SE-LSTM model achieved accuracies of 93.88%, 96.38%, and 95.33%, with the lowest overfitting values of 2.13%, 1.55%, and 1.32%, respectively. The SE-LSTM model demonstrated a balanced performance in both accuracy and stability, indicating superior capability in predicting MPs stress levels.

Although the predicting effect based on full wavelength has achieved satisfactory results, the large number of bands poses challenges for rapid detection. To enhance detection efficiency, characteristic wavelength selection was performed, and the feasibility of four algorithms, including successive projection algorithm (SPA), genetic algorithm (GA), particle swarm optimization (PSO), and bootstrapping soft shrinkage (BOSS), was evaluated. The characteristic wavelengths selected by different methods exhibited notable overlap ( Figure 8 ). For example, under all three types of microplastic stress, wavelengths such as 425 nm, 431 nm, 439 nm, and 550 nm were frequently identified, suggesting that MPs exposure influences pigment absorption characteristics in rice leaves (Zhao et al., 2023), particularly those related to chlorophyll and carotenoids. Regarding differences among the wavelength selection methods, SPA tended to select a smaller number of bands, which likely emphasized on minimizing spectral redundancy and maximizing band independence. In contrast, a substantially larger number of wavelengths converging VNIR region were selected by GA and PSO, which is likely due to their global search strategies for the identification of stress-related features. The BOSS algorithm showed a degree of stochasticity in band selection; however, the wavelengths selected overlapped with those from others. This suggests that BOSS balances the retention of critical spectral information with greater selection flexibility.

The accuracy of the calibration set, cross-validation set and validation set for stress concentration prediction with characteristic wavelengths were shown in Figure 9 . Compared to full-wavelength models, the multi-wavelength model significantly reduces the number of inputs and maintains a relatively high classification accuracy rate. The optimal prediction accuracies for PET, PS, and PVC were 85.7%, 90.3%, and 94.5% in the multi-wavelength detection model, respectively. Regarding the unsatisfactory prediction of PET, the accuracy rate of full-wavelength model is 93.88%, while that of multi-wavelength model is only 83.15%. In contrast, PVC prediction by two models almost achieved the same prediction results, with full-wavelength and multi-wavelength models achieving 95.33% and 94.51%, which was almost comparable. Among the wavelength selection methods, the GA achieved the optimal results for PET and PS, whereas the BOSS algorithm performed optimally for PVC, which can be used for subsequent analysis. These findings suggest that multi-wavelength models significantly reduced data dimensionality and preserving critical spectral information, enabling efficient and rapid detection.

To explore the importance of characteristic wavelengths and their corresponding biochemical associations under MPs stress, the SHAP analysis was employed to enhance model interpretability. As shown in Figure 10 , characteristic wavelengths contributing to stress detection in the spectral data were identified and their potential associated components were analyzed. The global importance of individual wavelengths was visualized in the bar chart ( Figure 10a ), and Figures 10b–d illustrates feature importance variations across stress levels (0, 10, and 100 mg/L) using bee swarm plots. From a global perspective, the top eight discriminative wavelengths for PET, PS, and PVC are as follows: PET: 445, 472, 536, 588, 623, 662, and 937 nm; PS: 425, 435, 461, 502, 687, 736, 912, and 940 nm; PVC: 425, 451, 488, 653, 689, 736, 760, and 939 nm. Notably, there is considerable overlap among the three groups in the 425–502 nm (blue light) range, potentially associated with chlorophyll a/b and carotenoid absorption; in the 687–760 nm (red-edge) region, likely related to chlorophyll fluorescence sensitivity; and in the 736–940 nm (near-infrared) region, which may correspond to water or cellulose overtone vibrations (He et al., 2024). These findings suggest that stress-induced disruptions in pigment metabolism, cellular structure, and water transport may serve as effective indicators for distinguishing stress levels. In terms of differences, PET detection showed unique contributions at 445 nm (a chlorophyll absorption peak in the blue-violet region) and 937 nm (a third overtone of O–H bond vibrations). For PS, a strong contribution was observed at 912 nm, corresponding to C–H stretching vibrations. Additionally, the feature at 760 nm exhibited heightened sensitivity for PVC, which attributed to the second overtone of O–H vibrations, indicating possible shifts in dominant physiological responses across different stress concentrations.

The bee swarm plots further revealed the concentration dependence of feature contributions: from top to bottom is the distribution of wavelength importance, and each point in the map represents the SHAP value of the sample at each wavelength. The distribution of SHAP values reflects the contribution level through color mapping, with red representing high values and blue representing low values. The positions of the positive and negative half-axes respectively represent the promoting or inhibitory effects of the features on the model prediction. In PET detection, the distribution of SHAP values in the carotenoid absorption valley at 472 nm and the flavonoid sensitive zone at 439 nm in the control group showed high synchrony, which might be related to the homeostasis of secondary metabolites. The strong negative contribution of the chlorophyll reflection peak at 540nm and the water stress marker at 937nm in the 100 mg/L group suggests that high concentrations of microplastics may intensify the stress response by damaging photosynthetic organs and intensifying transpiration. Meanwhile, there are differences in the single weight distribution of different types and concentrations of MPs. This may be related to the dynamic differentiation of the SHAP contribution mode and the nonlinear weight distribution of the model for capturing physiological markers at different stress stages.

The overall trends of the mid-infrared spectra in the vein and mesophyll regions of rice leaves under PET microplastic stress were similar ( Figure 11 ), but different spectral trends were presented in the functional group region (4000–2800 cm^-1) and the fingerprint region (1800–600 cm^-1). The absorption peak intensities of the veins and mesophyll in the control group were generally low. However, multiple characteristic peaks showed with PET concentration increased, indicating that MPs triggered dynamic responses of the molecular structure at the subcellular scale. In the vein spectra ( Figure 11a ), the peak intensity of the O-H/N-H stretching vibration at 3352 cm^-1 significantly increased with concentration increased, which might be related to the obstruction of water transport in the vascular bundle or the accumulation of woody fluid (Bunaciu et al., 2012). Combined with the PCA loading ( Figure 11b ), the synergistic enhancement of the peak at 1776 cm^-1 in PC₁ loading of leaf vein and the stretching vibration of C=O of the ester group at 1730 cm^-1 in original spectra points to the physical adsorption and chemical hydrolysis process of the PET bulk or its degradation products on the inner wall of the fiber tube (Mecozzi and Nisini, 2019). Furthermore, the stretching vibration of C-O at 1248cm^-1 in PC₂ loading and the synchronous increase of the 1244 cm^-1 peak in the original spectra indicate that PET induces enhanced lignification of the secondary wall of vascular bundle sheath cells through mechanical friction.

The spectral response in the mesophyll region ( Figure 11c ) was dominated by membrane system damage and dynamic regulation of the cell wall. The asymmetric stretching vibration of CH₂ of lipid at 2898 cm^-1 reflects the accumulation of malonaldehyde in chloroplasts (Brandl, 2013), which is consistent with chlorophyll scattering at 690 nm in VNIR spectra. The C-H bending vibration of pectin methylation at 1469 cm^-1 in PC₁ loading suggests that a decrease in cell wall rigidity to alleviate PET induced stress. The stretching vibration of C=O of the ester group at 1730 cm^-1 and the bending vibration of C-H of the benzene ring at 841 cm^-1 are related to PET. The peak intensity increases with the increase of concentration in the vein, while only weakly appears or is completely absent in the mesophyll. In conclusion, PET MPs are more likely to be transported through the vascular bundle system and enriched on the inner walls of the vessels, while there are fewer mesophyll cells, which may be related to the rejection of large-sized particles by exocytosis.

Due to the overlapping of multiple absorption peaks in SR-FTIR spectra and the unclear mechanism of biochemical changes in leaves. The 2DCOS technology was used to explore the biochemical change mechanism of leaf veins and leaves caused by external disturbances of rice under different concentrations of MPs. In addition, the absorption peaks in the functional group region (4000–2800 cm^-1) and the fingerprint region (1800–800 cm^-1) are of different complexities, so, they are explored in two regions. The automatic peak in the synchronous spectra represents the spectral intensity at the corresponding coordinate, while the asynchronous spectra represent the sequence of component peak intensity changes caused by external disturbances.

The 2DCOS spectra of the functional group regions of leaf veins under PET MPs stress were shown in Figures 12a, b . The two automatic peaks of 3433 cm^-1 and 2918 cm^-1 in the synchronous spectra correspond respectively to the stretching vibration of O-H/N-H and the asymmetric stretching vibration of CH₂, indicating that the change of water state and the disorder of membrane lipid metabolism are the response characteristics of leaf veins. The relationship of component changes was clarified from 3356 → 2910 → 2863→ 3169 → 3506 cm^-1 in the asynchronous spectra. The O-H vibration at 3356 cm^-1 reflected the interference of PET MPs on the water transport of vascular bundles. Subsequently, CH₂ at 2910 cm^-1 and 2863 cm^-1 marked the initiation of membrane lipid peroxidation, and N-H vibration at 3169 cm^-1 indicates the conformational change of membrane-binding proteins. The O-H vibration at 3506 cm^-1 may reflect the increase in the hydroxylation degree of polysaccharides in the cell wall, which is the adaptive response of vascular bundle tissues to persistent stress.

In the synchronous spectra of the leaf vein fingerprint region ( Figure 12c ), there are automatic peaks such as C-H curvature of the benzene ring of PET at 840 cm^-1, the C-O of lignin at 1258cm^-1, and the C=O of ester group of PET at 1724 cm^-1. The accumulation of MPs in the vascular bundle and the defense responses it triggers are demonstrated. The asynchronous spectra ( Figure 12d ) revealed the dynamic change processes of the cell wall components and metabolites (1115→1439→1373→1252→1648 →975 cm^-1). Firstly, the cellulose vibration at 1115 cm^-1 indicates the remodeling of polysaccharides in the cell wall. Subsequently, the methylation of pectin at 1439 cm^-1 further regulates the performance of the cell wall. Then, the lignin-related vibrations at 1252 cm^-1 revealed the strengthening process of the secondary cell wall. Finally, the protein denaturation at 1648 cm^-1 and the PET benzene ring characteristics at 975 cm^-1 suggested the enzyme function impairment and the accumulation of MPs degradation products caused by continuous stress.

The responses of mesophyll tissue in the functional group regions ( Figures 13a, b ) differ from those in the veins. The automatic peaks of O-H stretching at 3400 cm^-1 and CH₂ stretching at 2880 cm^-1 in the synchronous spectrum indicate that the PET stress of mesophyll cells is mainly concentrated on the cell wall and membrane system. The sequence of peak changes in the asynchronous spectrum (3443→3187→3359→2887→3839 cm^-1) reveals the response path: Firstly, it causes the rapid adjustment of free hydroxyl groups in the polysaccharide at 3443 cm^-1. Secondly, it leads to the vibration of N-H of protein at 3187 cm^-1, indicating that MPs causes cell wall damage in mesophyll tissue. Then, the vibration of membrane lipids at 2887cm^-1 and hydroxyl groups at 3839 cm^-1 led to the damage of mesophyll cells under continuous stress, which contrasted with the vein tissue.

In the synchronous spectra ( Figure 13c ) of the mesophyll fingerprint region, the auto-peaks such as C-O-C of cellulose at 1040 cm^-1, CH₂ bending of pectin at 1463 cm^-1, and C=O of ester group at 1730 cm^-1 were associated with concentration dependence, reflecting the effect of PET stress on the mesophyll cell wall. The sequence of component changes in the asynchronous spectra ( Figure 13d ) is 1226→1087→1020→968→1158→1337→1738→1647→1463 cm^-1. Firstly, the vibrations of lignin/phenolic acid at 1226 cm^-1 and the cellulose at 1087 cm^-1 indicate that secondary metabolic activation and cell wall enhancement are the initial defense strategies. Then, 1020 cm^-1 and 968 cm^-1 suggested the remodeling of the carbon metabolism pathway, which might be a response to insufficient energy supply. Finally, the vibration of ester group at 1738 cm^-1 and the protein denaturation at 1647 cm^-1 revealed the cumulative damage of PET chemical toxicity to mesophyll cells.