Exploring Drought Response: Machine-Learning-Based Classification of Rice Tolerance Using Root and Physiological Traits

Abstract

Drought is a key limitation for rice productivity. While oxidative stress markers like hydrogen peroxide (H₂O₂) are important for drought adaptation, the predictive value of combining root anatomical and physiological traits is underexplored. We assessed 20 rice cultivars under drought and control conditions using a random forest, a multi-layer perceptron, and a SHAP-optimized stacking ensemble. The stacking ensemble achieved the highest classification accuracy (81.8%) and identified hydrogen peroxide, relative water content, and endodermis inner circumference as key predictors. SHAP analysis revealed important interactions between root anatomical and physiological traits, providing new biological insights into drought tolerance. Our integrative approach, supported by robust cross-validation, improves predictive power and transparency for breeding drought-resilient rice cultivars.

1. Introduction

Drought stress is one of the most severe abiotic stresses affecting global agricultural productivity, particularly for staple crops such as rice (Oryza sativa L.), which is highly sensitive to water limitations [1,2]. Water deficit conditions disrupt cellular homeostasis, leading to impaired photosynthesis, reduced stomatal conductance, and an imbalance in oxidative processes [2,3,4]. A primary response to drought is the decline in relative water content (RWC) [5,6], along with the accumulation of reactive oxygen species (ROS) that cause oxidative damage, as evidenced by elevated malondialdehyde (MDA) and hydrogen peroxide (H₂O₂) [7,8,9] and by chlorophyll degradation, which limits photosynthetic efficiency and growth [10,11].

The urgent need to develop drought-tolerant rice cultivars necessitates efficient and accurate screening methods that integrate the physiological and anatomical traits associated with drought adaptation [12,13]. Traditional drought tolerance assessment methods, such as phenotypic selection, physiological measurements, and molecular marker-assisted selection (MAS), have limitations in capturing the complexity of drought tolerance [14,15]. Phenotypic selection often lacks the resolution needed to capture the drought-induced physiological changes [13,16]. To overcome these limitations, machine learning (ML) and artificial intelligence (AI) have emerged as transformative tools in plant science, enabling the integration of large-scale physiological, anatomical, and omics datasets to enhance predictive accuracy [17]. ML approaches excel in capturing non-linear interactions, identifying key features, and enhancing classification efficiency in drought stress phenotyping [18,19,20].

Recent advancements in AI-driven plant phenotyping underscore the potential of ML for drought tolerance prediction, trait association studies, and multi-omics integration [21,22,23]. Specifically, random forest (RF) models have been widely employed for feature selection and ranking, providing insights into the most influential physiological traits that drive drought adaptation [24,25]. Meanwhile, deep-learning architectures, including multi-layer perceptron (MLP) networks, have been utilized to model complex trait interactions and enhance classification precision in plant stress phenotyping [26,27]. However, most existing ML studies focus primarily on physiological data, often neglecting the critical role of root anatomical traits in drought resistance [28,29,30]. To address this gap, an integrated stacking ensemble ML framework that combines physiological and anatomical features is needed to improve drought tolerance classification accuracy and model interpretability [31,32].

While ML-based drought tolerance prediction has made significant progress, several research gaps remain. Existing models predominantly rely on physiological indices, overlooking the root anatomical parameters that influence water uptake efficiency [33,34]. Additionally, while ANN models offer high accuracy, they often function as “black boxes,” making a biological interpretation of feature importance challenging [35,36]. Current studies rarely explore stacking ensembles that integrate multiple ML classifiers to enhance drought tolerance classification [37,38]. This study aims to bridge these gaps by developing a novel stacking ensemble ML framework that integrates a random forest, neural networks, and a stacking ensemble model to classify drought tolerance based on both physiological and anatomical traits. By leveraging SHAP (Shapley additive explanations) for feature interpretability, this research aims to provide high predictive accuracy and valuable biological insights into drought adaptation mechanisms [37,38,39].

Thus, our study introduces a robust, interpretable, and reproducible AI-driven pipeline for rice drought tolerance prediction by integrating both physiological and root anatomical traits within a novel stacking ensemble machine-learning framework optimized with SHAP. This approach not only improves classification accuracy but also provides transparent insight into synergistic trait interactions under drought stress, advancing precision phenotyping in rice. Our work is novel in that it unifies diverse phenotypic data, addresses both model performance and interpretability, and delivers a validated workflow that supports the acceleration of drought-resilient rice breeding under climate change.

2. Materials and Methods

2.1. Plant Material

For rice (Oryza sativa L.), 20 landrace varieties and 2 comparative varieties (i.e., KDML105 as a susceptibility and CT9993 as a tolerance to drought conditions) were selected for this study, as shown in Appendix A 

Table A1. List of rice cultivars used in the study, their drought tolerance classification, and cultivar type.

Cultivar	Drought Tolerant Class	Type
Kab yang	tolerant	Landrace
Kam	tolerant	Landrace
Khao soi	moderate	Landrace
Luang bunma	moderate	Landrace
Luang kaeo	susceptible	Landrace
Hang nak	tolerant	Landrace
Hom thung	moderate	Landrace
Long ma	moderate	Landrace
Samart	susceptible	Landrace
Sam ruang	moderate	Landrace
Mae phueng	moderate	Landrace
Mayom	moderate	Landrace
Mafai	moderate	Landrace
Maprang	moderate	Landrace
Dor yuan	susceptible	Landrace
Jaew daeng	moderate	Landrace
Khithom dam	susceptible	Landrace
Phra in	tolerant	Landrace
I-daeng noi	moderate	Landrace
I-khiao non thung	tolerant	Landrace
CT9993	tolerant	Tolerant check (CT9993)
KDMLl105	moderate	Susceptible check (KDML105)

. Seeds were soaked using a sterilization method of 0.5% sodium hypochlorite for 10 min, then rinsed thoroughly with distilled water for 1 min 3 times. The sterilized seeds were sown in trays containing moisture germination paper tower and germinated under controlled conditions. After 3 days, uniform seedlings were transplanted into a tray containing vermiculite planting media and maintained under these growth conditions until the initiation of drought treatments.

2.2. Experimental Design

Greenhouse experiments were conducted. The two main treatments were well-watered (control), and PEG 6000 simulated drought. In the well-watered treatment, plants were cultivated in Yoshida nutrient solution [40] at full nutrient and water availability. Plants were cultivated in Yoshida nutritional solution with PEG 6000 to simulate drought. Dissolving 20% PEG 6000 in Yoshida nutrition solution and mixing carefully to ensure complete dissolution. Treatments were given 30 days after transplanting, and the drought lasted 10 days. To ensure nutrient availability, nutrient solutions were refreshed every 4 days. Three replicates of 10 plants each were grown for each treatment.

2.3. Physiological Measurements

2.3.1. Proline Content

Proline levels were measured using Bates et al.’s [41] technique. In 10 mL of 3% sulfosalicylic acid, 0.5 g of fresh leaf tissue were homogenized and centrifuged at 3000 rpm for 10 min. Acid ninhydrin and glacial acetic acid were mixed 1:1:1 in the supernatant and heated at 100 °C for an hour. The process was stopped by cooling the mixture in an ice bath. After adding 4 mL of toluene, the chromophore was extracted. We then examined the toluene layer’s absorbance at 520 nm with a spectrophotometer.

2.3.2. Chlorophyll Content

Chlorophyll content was measured using the method of Arnon (1949). Leaf samples (0.2 g) were homogenized in 80% acetone and centrifuged at 5000 rpm for 5 min. The supernatant was collected, and the absorbance was measured at 645 and 663 nm. Chlorophyll a, chlorophyll b, and total chlorophyll concentrations were calculated using the following equations:

Chlorophyll a = 12.7(A663) − 2.69(A645)

(1)

Chlorophyll b = 22.9(A645) − 4.68(A663)

(2)

Total chlorophyll = 20.2(A645) + 8.02(A663)

(3)

2.3.3. Malondialdehyde (MDA) Content

MDA content, an indicator of lipid peroxidation, was determined using the thiobarbituric acid (TBA) assay [42]. Fresh leaf tissue (0.5 g) was homogenized in 5 mL of 0.1% trichloroacetic acid (TCA) and centrifuged at 10,000 rpm for 10 min. The supernatant was mixed with 4 mL of 20% TCA containing 0.5% TBA and heated at 95 °C for 30 min. After cooling, the mixture was centrifuged at 10,000 rpm for 10 min, and the absorbance was measured at 532 and 600 nm. MDA content was calculated using the extinction coefficient of 155 mM⁻¹ cm⁻¹.

2.3.4. Relative Water Content (RWC)

RWC was measured using the method described by Turner (1986). Fresh leaf samples were weighed (fresh weight, FW), then rehydrated in distilled water for 4 h to obtain the turgid weight (TW). Samples were then oven-dried at 60 °C for 48 h to obtain the dry weight (DW). RWC was calculated using the formula:

RWC(%) = (TW − DW/FW − DW) × 100

(4)

2.4. Root Anatomical Measurements

Root samples were collected at the end of the drought treatment period. Roots were carefully washed to remove planted media particles and fixed in FAA solution (formalin–acetic acid–alcohol) until used for cross-section. Fixed root samples were dehydrated through a graded ethanol series and sectioned using a free-hand section method. Transverse sections (10–15 µm thick) were stained with safranin O and observed under a light microscope (Axio Imager 2; Carl Zeiss AG, Germany) and analyzed via Zen 3.5 software (Carl Zeiss AG, Germany).

2.4.1. Cortex Area

The cortex area was measured using image analysis software. Microscopic images of root sections were captured, and the cortex area was delineated and quantified.

2.4.2. Endodermis Characteristics

Endodermis thickness was measured at four equidistant points around the root circumference using a calibrated ocular micrometer.

2.4.3. Exodermis Characteristics

Exodermis characteristics, such as the number of cell layers and lignification, were assessed qualitatively and quantitatively using light microscopy and image analysis.

2.4.4. Metaxylem Vessel Dimensions

Metaxylem vessel diameter and number were measured in transverse root sections using image analysis software. The average vessel diameter was calculated from measurements of at least 10 vessels per root.

2.5. Data Summary and Descriptive Statistics

Before applying machine-learning models, a descriptive statistical analysis was conducted to summarize the dataset structure and ensure data quality. The dataset comprised physiological and anatomical traits measured under well-watered and drought-stressed conditions, with a total of 132 observations. Normality of the dataset was assessed using the Shapiro–Wilk test, and homogeneity of variance was verified using Levene’s test. Pearson’s correlation analysis revealed significant relationships between RWC and oxidative stress markers (MDA and H₂O₂), indicating drought-induced physiological stress. Boxplots and histograms were used to visualize data distribution and detect potential outliers. All statistical analyses were conducted using Python (version 3.12.6) within a Jupyter Notebook environment (version 7.2.2). The following Python libraries were used:

Pandas (2.2.3) [43]: for data manipulation and summary statistics, numpy (2.1.0) [44]: for numerical computations, scipy.stats (1.14.1) [45]: for hypothesis testing, including normality and variance tests, matplotlib (3.10.0) [46] and seaborn (0.13.2) [47]: for data visualization (histograms, boxplots, and correlation heatmaps), and statsmodels (0.14.5) [48]: for additional statistical analysis, including regression modeling. This statistical foundation ensures the dataset’s integrity before machine learning applications and strengthens the interpretation of results.

2.6. Machine Learning and AI Implementation

The integration of machine-learning (ML) and artificial intelligence (AI) techniques in plant stress research enhances predictive capacity and provides novel insights into physiological and anatomical responses. This study leverages Python (version 3.12.6) [49] within a Jupyter Notebook environment [50], operating on an HP 15.6-inch Laptop PC 15-d5000 equipped with a 12th Gen Intel^® Core™ i7-1260P processor (2.10 GHz, 16 GB RAM) and Windows 11 (64-bit). The analysis was conducted using well-established machine learning libraries, including scikit-learn (1.6.1) [51] (Pedregosa et al., 2011), TensorFlow (2.18.0)/Keras (3.8.0) [52], XGBoost (2.1.4) [53], and SHAP (0.46.0) for explainability [39].

2.6.1. Data Overview and Target Variable

The dataset comprised physiological and root anatomical traits of rice cultivars under both well-watered and drought conditions. The target variable for classification was drought tolerance level, categorized into tolerant, moderate, and susceptible groups. The features included relative water content (RWC), hydrogen peroxide (H₂O₂), malondialdehyde (MDA), chlorophyll content (total, a, and b), plant height, endodermis inner circumference, exodermis thickness, cortex area, and vascular cylinder.

To mitigate class imbalance, SMOTE (synthetic minority oversampling technique) [54] was applied, ensuring equal representation of all drought tolerance groups. Data preprocessing involved StandardScaler normalization [51], and an 80–20% train–test split was employed with stratified sampling to preserve class distributions.

2.6.2. Random Forest Model

Random forest, an ensemble-learning method, was chosen due to its ability to manage high-dimensional data, handle feature interactions, and provide insights through feature importance analysis. This model constructs multiple decision trees and aggregates their predictions to enhance classification stability. To determine the key drought tolerance predictors, Gini impurity-based ranking and permutation importance [55] were applied.

2.6.3. Neural Network Model

A multi-layer perceptron (MLP) [56] neural network was implemented to capture the non-linear relationships between physiological and anatomical traits. The architecture consisted of three hidden layers with BatchNormalization [57] and LeakyReLU [58] activation (α = 0.1) to enhance convergence. Dropout regularization [59] (30%) was included to mitigate overfitting. The Adam optimizer [60] was used for training, with early stopping (patience = 10) and adaptive-learning rate reduction (factor = 0.5, min_lr = 1 × 10⁻⁵). Feature importance was evaluated using SHAP analysis [39] to provide interpretable model outputs.

2.6.4. Stacking Ensemble Model

A stacking ensemble approach [61] was developed to integrate multiple classifiers for improved predictive performance. The base models included random forest (RF) [55], XGBoost (XGB) [53], support vector machine (SVM) [62], and multi-layer perceptron (MLP) [56], with XGBoost [53] serving as the meta-learner. Predictions from each base model were combined as meta-features for the final classification. A SHAP [39] analysis was performed to highlight the most influential physiological and anatomical traits affecting drought tolerance.

To assess and compare model performance, confusion matrices, receiver operating characteristic (ROC) curves [63], model classification reports, and feature importance rankings were used. These evaluations provided insights into model reliability, predictive power, and the key factors influencing classification accuracy.

2.6.5. Reproducibility and Computational Environment

Reproducibility is a crucial aspect of ML-based plant science research. All computations were executed in Python (version 3.12.6) [49] using Jupyter Notebook [50] to ensure traceability. The following libraries were utilized: scikit-learn [51] for ML model implementation and evaluation, TensorFlow/Keras [52] for deep-learning-based classification, XGBoost [53] for gradient boosting techniques, and SHAP [39] for model interpretability and feature ranking. This methodological framework facilitates the adoption of ML and AI in plant stress data analysis, offering a reproducible and interpretable approach for identifying key drought tolerance traits. To enhance clarity and reproducibility, we provide both schematic block diagrams and detailed algorithmic steps for each machine-learning model applied in this study (Figure 1). This includes random forest (RF), multi-layer perceptron (MLP), and the stacking ensemble. The full workflow encompasses data preprocessing, class balancing with SMOTE, train–test splitting, model training, feature importance analysis, and evaluation using ten classification metrics.

In this study, a multi-model machine-learning strategy was adopted to both maximize prediction accuracy and extract interpretable biological insights into rice drought tolerance. Each model was selected to address distinct analytical objectives, as summarized in

Table 1. Overview of machine-learning models, rationale, and unique contributions as implemented in this study.

Model	Rationale and Role	Unique Contribution and Core Algorithmic Summary
Random Forest (RF)	Handles high-dimensional, mixed-type data; interpretable feature ranking	Establishes baseline accuracy and trait importance; uses bootstrap aggregation, Gini impurity minimization, and feature importance via mean decrease in Gini.
Multi-Layer Perceptron (MLP)	Captures complex, nonlinear relationships	Detects subtle interactions among traits; fully connected layers with LeakyReLU activation and softmax output, optimized by Adam and categorical cross-entropy loss.
Stacking Ensemble	Integrates diverse classifiers for improved accuracy and robustness	Maximizes performance by combining RF, SVM, XGBoost, and MLP base models, with XGBoost as the meta-learner to determine the final classification.

. This multi-model framework was designed so that each algorithm complements the others: random forest provides interpretability and robust feature ranking, MLP uncovers complex nonlinear patterns, and the stacking ensemble achieves maximum predictive power by integrating multiple models. This approach enables rigorous benchmarking of model performance, as well as in-depth biological interpretation. Importantly, SHAP analysis was applied throughout to provide transparent, quantitative insights into feature contributions and trait interactions, thereby offering actionable guidance for breeders and plant scientists in rice drought phenotyping

The explicit algorithmic steps and corresponding mathematical equations are provided in the Supplementary Materials.

2.6.6. Model Performance Evaluation and Statistical Comparison

To ensure both practical and statistical rigor, model performance was evaluated using two approaches. Reporting the best single train–test split facilitates direct comparison with prior work and highlights the model’s peak potential. Five-fold cross-validation provides a more robust estimate of generalizability and model stability. For each approach, ten standard classification metrics were calculated: accuracy, balanced accuracy, macro precision, macro recall, macro F1-score, ROC AUC, log loss, Matthews correlation coefficient, Cohen’s kappa, and hamming loss. The mean and standard deviation across folds are reported for cross-validation.

Statistical comparisons between models for all metrics were conducted using paired t-tests or Wilcoxon signed-rank tests.

The authors acknowledge the use of ChatGPT-4o (OpenAI, version as of June 2025) as a supportive instrument during the production of this manuscript. In particular, ChatGPT-4o was employed to aid in the formulation and refinement of the methodology section, as well as the generation of initial Python code templates for machine-learning techniques, including SMOTE, random forests, multi-layer perceptrons (MLP), stacking ensembles, and SHAP analysis. The authors thoroughly reviewed, tested, modified, and executed all code and text outputs. The authors are solely responsible for the accuracy and integrity of the research presented, having conducted the study design, code implementation, data analysis, and interpretation.

3. Results

3.1. Trait Variability and Physiological Correlations Under Contrasting Water Regimes

A descriptive statistical analysis was conducted to characterize the distribution, variability, and interrelationships among the measured physiological and anatomical traits under both well-watered and drought-stressed conditions. The dataset comprised 132 observations across 20 rice landrace cultivars and two reference checks, representing a diverse phenotypic response spectrum. Key traits included plant height, relative water content (RWC), oxidative stress markers (malondialdehyde [MDA] and hydrogen peroxide [H₂O₂]), chlorophyll pigments (chlorophyll a, b, and total), and detailed root anatomical metrics. Normality assessment using the Shapiro–Wilk test indicated that several physiological traits, particularly RWC, MDA, and H₂O₂, significantly deviated from a Gaussian distribution (p < 0.05), reflecting heterogeneous stress responses across genotypes. Levene’s test further revealed heteroscedasticity between water treatments in RWC, MDA, and chlorophyll content, underscoring drought-induced physiological changes. Pearson’s correlation analysis revealed strong positive relationships between RWC and chlorophyll traits, as well as between MDA and H₂O₂, highlighting their co-regulated behaviors as indicators of oxidative stress. The comprehensive trait interactions are visualized in Appendix A, Figure A2, which displays a Pearson correlation heatmap illustrating patterns of association among 16 traits under water-stress conditions. Additionally, treatment-specific variability in key anatomical and physiological traits was visualized using histograms overlaid with kernel density estimations and boxplots (Appendix A, Figure A1). These visualizations highlight treatment-driven shifts, trait-specific skewness, and the presence of genotype outliers, offering deeper insight into stress adaptation mechanisms. Together, these data provide a statistical and visual foundation for subsequent machine-learning analyses.

3.2. Random Forest Analysis

A random forest (RF) classification model was developed to predict drought response categories, namely tolerant, moderate, and susceptible, among diverse rice cultivars, based on a comprehensive set of physiological and root anatomical traits. After preprocessing and class rebalancing using the synthetic minority over-sampling technique (SMOTE), the optimized model achieved an overall classification accuracy of 77%, as evaluated on an independent test set. The model’s internal feature ranking, visualized in Figure 2a, identified hydrogen peroxide (H₂O₂) as the most influential trait, followed by vascular cylinder thickness, plant height, endodermis thickness, and cortex size. These traits collectively reflect a combination of oxidative stress signaling and root morphological adaptation, aligning with known mechanisms of drought tolerance. In contrast, features such as chlorophyll b, malondialdehyde (MDA), and external endodermal circumference contributed minimally to prediction and were subsequently excluded during model refinement. The confusion matrix (Figure 2b) revealed that the model achieved perfect classification of susceptible cultivars, high performance in the tolerant group, and minor misclassification in the moderate class, likely due to phenotypic overlap. Complementary to this, the multiclass ROC curves (Figure 2c) demonstrated excellent class separability, with AUC values of 0.99 for both the tolerant and susceptible groups and 0.94 for the moderate group. These results confirm the model’s high sensitivity and specificity, particularly in identifying extreme phenotypes under drought stress.

Collectively, these findings underscore the utility of random forest models for drought response classification in rice and highlight interpretable, biologically relevant traits that can support future phenotypic screening and breeding strategies under water-limited conditions.

3.3. Neural Network Model Analysis

The multi-layer perceptron (MLP) neural network exhibited strong performance in classifying rice genotypes into drought tolerance categories, namely moderate, susceptible, and tolerant, based on physiological and root anatomical features. As illustrated in the confusion matrix (Figure 3a), the model accurately identified most instances of the susceptible and tolerant classes, while a degree of overlap was observed in the moderate group, which is biologically plausible given its intermediate phenotype. This trend was further confirmed by the ROC curves (Figure 3b), which revealed high discriminative ability with area under the curve (AUC) values of 0.98 for the susceptible class, 0.95 for the tolerant, and 0.91 for the moderate. These results underscore the model’s capacity to capture complex, non-linear patterns in the input data and reflect its strength in multi-class classification under physiological heterogeneity.

To gain insight into the internal decision-making process of the MLP, we applied SHAP (Shapley additive explanations) and permutation importance techniques (Figure 3c,d). Both analyses consistently highlighted hydrogen peroxide (H₂O₂), plant height, and endodermis inner circumference as the most influential features driving the model’s predictions—an alignment that reinforces their biological relevance in drought stress response. SHAP summary plots revealed that higher values of these traits increased the likelihood of a genotype being classified as susceptible, suggesting a physiological profile marked by oxidative stress and elongation. Permutation importance rankings confirmed this pattern, with H₂O₂ emerging as the top predictor. The architecture of the MLP model (Figure 3e) consisted of ten input nodes connected to three hidden layers of 512, 256, and 128 neurons, followed by a three-node softmax output layer. This design allowed the model to learn intricate interactions among input traits, enabling accurate classification while maintaining interpretability through post hoc explanation methods. These findings collectively demonstrate the effectiveness of neural network-based modeling in elucidating the physiological and anatomical signatures of drought tolerance in rice.

3.4. Stacking Ensemble Model Analysis

To enhance classification performance and model explainability, we implemented a stacking ensemble strategy that integrated four base classifiers, namely random forest (RF), XGBoost (XGB), support vector machine (SVM), and multi-layer perceptron (MLP), into a two-layer stacking framework with logistic regression as the meta-learner. While the initial ensemble performance stabilized at an accuracy of 0.7727, further optimization was achieved by incorporating SHAP-based feature selection. Using SHAP (Shapley additive explanations), we quantified the contribution of each input trait and subsequently retrained all base models using the top 10 most influential features. This refined stacking ensemble model reached a final test accuracy of 0.8181, representing the best-performing configuration in our study. The SHAP interaction plot (Figure 4a) reveals important nonlinear interdependencies among key physiological traits, particularly between cortex, chla, and chlb. These interactions suggest that drought responsiveness cannot be fully captured by individual traits in isolation, highlighting the need for models that accommodate synergistic physiological effects.

Classification performance across three drought-responsiveness classes (susceptible, moderate, and tolerant) was evaluated via a confusion matrix analysis (Figure 4b). The stacking ensemble model correctly identified tolerant genotypes with high reliability (recall = 0.93), while maintaining balanced precision across other classes. A corresponding ROC analysis (Figure 4c) demonstrated strong overall discriminative capacity, with area under the curve (AUC) values of 0.91, 0.78, and 0.90 for the susceptible, moderate, and tolerant groups, respectively. The SHAP-derived importance rankings (Figure 3d) consistently prioritized traits linked to oxidative stress response and anatomical resilience. Physiological markers such as relative water content (RWC) and hydrogen peroxide (H₂O₂) emerged as the most influential, followed by root anatomical metrics, including endo_in_circ, vascul_thick, and cortex. These results suggest that both hydraulic maintenance under water stress and root structural adaptations play pivotal roles in conferring drought tolerance—particularly among landrace cultivars with intermediate phenotypes.

3.5. Model Comparison and Statistical Validation

To comprehensively assess model performance, we compared random forest (RF), multi-layer perceptron (MLP), and the stacking ensemble across ten standard metrics using both the best single train–test split and five-fold cross-validation (

Table 2. Mean (±SD) across 5 folds and best single-split for each model.

Metric	5 Folds Cross Validation (Mean ± SD)			Best Single Split
	Random Forest	MLP	Stacking Ensemble	(Model/Value)
Accuracy	0.546 ± 0.067	0.485 ± 0.080	0.433 ± 0.100	Stacking Ensemble/0.818
Balanced Accuracy	0.469 ± 0.054	0.389 ± 0.066	0.320 ± 0.090	Stacking Ensemble/0.821
Precision (macro)	0.494 ± 0.052	0.403 ± 0.121	0.306 ± 0.109	Stacking Ensemble/0.830
Recall (macro)	0.469 ± 0.054	0.389 ± 0.066	0.320 ± 0.090	Stacking Ensemble/0.813
F1 Score (macro)	0.459 ± 0.058	0.380 ± 0.069	0.304 ± 0.090	Stacking Ensemble/0.812
ROC AUC (micro/ovo)	0.730 ± 0.033	0.702 ± 0.066	0.440 ± 0.066	Random Forest/0.973
Log Loss †	0.946 ± 0.032	2.188 ± 0.561	1.704 ± 0.269	MLP/0.555 (lowest)
Matthews Corrcoef	0.235 ± 0.073	0.086 ± 0.120	−0.033 ± 0.162	Stacking Ensemble/0.738
Cohen’s Kappa	0.222 ± 0.081	0.079 ± 0.120	−0.028 ± 0.155	Stacking Ensemble/0.728
Hamming Loss †	0.454 ± 0.077	0.515 ± 0.081	0.567 ± 0.093	Stacking Ensemble/0.182 (lowest)

Note: Mean and standard deviation (SD) are from 5-fold cross-validation for each model. “Best Single Split” is the highest (or lowest, for log loss and hamming loss) value achieved in any single split. ^†: For log loss and hamming loss, lower is better.

). Random forest demonstrated the highest mean values across most metrics in five-fold cross-validation, including accuracy (0.546 ± 0.067), balanced accuracy (0.469 ± 0.054), macro precision (0.494 ± 0.052), macro recall (0.469 ± 0.054), and macro F1-score (0.459 ± 0.058). RF also had the highest mean ROC AUC (0.730 ± 0.033) and the lowest mean log loss (0.946 ± 0.032) and hamming loss (0.454 ± 0.077), indicating strong and consistent multi-class classification. The stacking ensemble achieved the highest single-split performance in terms of accuracy (0.818), balanced accuracy (0.821), macro precision (0.830), and macro F1-score (0.812), illustrating its potential for optimal classification when trained and tested on a specific data split. However, its cross-validated mean metrics were generally lower than RF, with greater variability (accuracy 0.433 ± 0.100, hamming loss 0.567 ± 0.093). The MLP model performed moderately across all metrics, with cross-validated means between those of the RF and stacking ensemble for some metrics. Notably, MLP had the lowest single-split log loss (0.555), reflecting fewer high-confidence errors under optimal conditions. Statistical comparisons (

Table 3. Statistical comparison of model performance across 10 metrics (p-values).

Metric	RF VS. MLP	RF VS. Stacking Ensemble	MLP VS. Stacking Ensemble
Accuracy	0.098	0.063	0.273
Balanced Accuracy	0.074	0.017 *	0.273
Precision (macro)	0.120	0.007 *	0.138
Recall (macro)	0.074	0.017 *	0.273
F1 Score (macro)	0.137	0.007 *	0.206
ROC AUC (ovo, micro)	0.512	0.002 *	0.002 *
Log Loss †	0.007 *	0.006 *	0.045 *
Matthews Corrcoef	0.063	0.021 *	0.241
Cohen’s Kappa	0.083	0.021 *	0.214
Hamming Loss †	0.109	0.063	0.214

Note: (*) indicates statistical significance at p < 0.05 (paired t-test or Wilcoxon test). Dagger (^†) indicates log loss and hamming loss; lower values are better.

) revealed that the differences between RF and stacking ensemble, and between RF and MLP, were statistically significant for balanced accuracy, precision, F1-score, ROC AUC, log loss, MCC, and kappa (p < 0.05 in most cases), consistently favoring random forest for generalizability. Stacking ensemble’s higher best single-split metrics suggest it may be more sensitive to specific data partitions.

4. Discussion

The integration of physiological and root anatomical traits in this study provided new insights into the complex mechanisms underlying drought tolerance in rice. Across all three machine learning (ML) approaches, namely random forest (RF), multi-layer perceptron (MLP), and a SHAP-optimized stacking ensemble, the consistent emergence of oxidative stress markers (notably hydrogen peroxide, H₂O₂), root anatomical structures (such as endodermis inner circumference and vascular cylinder thickness), and classical physiological indicators (like relative water content, RWC, and plant height) as top-ranked features underscores their central role in mediating adaptive responses to water deficit.

Hydrogen peroxide (H₂O₂), identified as the most influential trait across models, plays a dual role in drought stress: as a damaging reactive oxygen species (ROS) and as a signaling molecule involved in stress-responsive gene expression [7,8,9]. Under drought, ROS accumulation—especially H₂O₂—is commonly observed due to enhanced photorespiration and disrupted electron transport chains in chloroplasts [3,7]. Our results reinforce previous findings that oxidative stress markers, particularly H₂O₂, are reliable indicators of drought sensitivity and can serve as early biomarkers for genotype screening [8,20].

Relative water content (RWC) emerged as a strong physiological predictor, reflecting the plant’s ability to maintain cellular hydration during drought—a classic trait associated with drought avoidance mechanisms [5,6,16]. Its significant correlation with chlorophyll content observed in this study is consistent with established physiological theories where RWC preservation helps maintain chloroplast integrity and delay senescence [3,11,16].

The anatomical traits, especially endodermis inner circumference and vascular cylinder thickness, provided mechanistic insights into water uptake and retention strategies. These structures are central to regulating radial water flow and limiting uncontrolled water loss through apoplastic pathways [30,33]. A thicker endodermis is associated with enhanced suberization and Casparian strip formation, which increase root selectivity and protect against desiccation [28,30]. Similarly, a robust vascular cylinder facilitates efficient axial water transport, improving water use under limited availability—a trait linked to improved drought adaptation in maize and other cereals [28,33].

The cortex area, often overlooked in drought phenotyping, showed strong interactions with photosynthetic traits such as chlorophyll a and b in the SHAP interaction analyses. Recent reports suggest that the development of root cortical aerenchyma and reduced metabolic activity in the cortex can enhance drought adaptation by lowering root respiration and facilitating deeper rooting [28,34]. These results align with the growing emphasis on integrating root anatomical efficiency into drought-tolerant ideotypes [28,30].

Interestingly, plant height, a trait traditionally associated with agronomic performance, also surfaced as a key predictor in both the RF and MLP models. While taller genotypes may have higher water demand, the association of height with drought tolerance could reflect favorable root-to-shoot signaling or deeper rooting capacity, particularly in landraces with balanced growth and stress adaptation [6,13,16]. These findings warrant further exploration into the physiological trade-offs between vegetative growth and stress resilience.

The stacking ensemble model, achieving the highest classification accuracy (81.81%), illustrated the power of combining diverse algorithms—each capturing distinct trait interaction patterns. Importantly, the SHAP-based feature refinement retained only biologically meaningful predictors, improving both interpretability and model performance. This layered approach supports earlier findings that ensemble learning, particularly stacking, enhances classification reliability in complex phenotypic datasets [37,38].

Beyond single train–test split evaluations, our use of five-fold cross-validation with ten complementary metrics provided a rigorous assessment of model generalizability and reliability. Random forest consistently demonstrated the highest mean performance across accuracy, balanced accuracy, precision, recall, F1-score, and ROC AUC, indicating its robustness for classifying drought tolerance under diverse data splits. The stacking ensemble, while achieving the highest peak accuracy in the best split, showed greater variability across folds—highlighting its sensitivity to data partitioning, but also its potential for optimal classification under certain scenarios. Importantly, the best single split remains valuable from a biological perspective, as it may represent the optimal conditions or data structure encountered in specific breeding programs or field environments. Nevertheless, single split results can overestimate performance due to chance data partitioning, while cross-validation provides a more balanced estimate of model generalizability—though still limited by sample size and experimental scope. Statistical comparisons revealed that the differences between models were significant for several key metrics, affirming the value of thorough multi-metric validation in phenotyping studies. Collectively, these results underscore the necessity of integrating both single split and cross-validated, multi-metric benchmarking to draw reliable and biologically meaningful conclusions in plant science machine-learning applications.

This study validates known mechanisms of drought tolerance and uncovers the predictive power of lesser-studied root traits, such as endodermis inner circumference. These traits can be incorporated into high-throughput phenotyping platforms using image-based anatomical screening, thus expanding the breeder’s toolbox beyond conventional physiological metrics [22,30,31]. Furthermore, the dominance of oxidative stress indicators like H₂O₂ across all models suggests that redox homeostasis remains a pivotal aspect of early drought stress detection and may be amenable to molecular marker-assisted selection [8,20].

Unlike previous studies that have focused exclusively on either physiological or anatomical parameters for drought response prediction [5,13], our work integrates both trait domains within a unified predictive framework. While most existing ML-based phenotyping relies on single-model classifiers or simple data fusion, we introduce a SHAP-optimized stacking ensemble that not only enhances classification accuracy but also provides interpretable insights into how trait combinations interact to influence drought tolerance [3,6]. This approach moves beyond conventional feature ranking to uncover biologically meaningful synergies between physiological and root anatomical adaptations—a dimension rarely explored in the plant phenotyping literature. By openly sharing our code and dataset, we further promote transparency and reproducibility for the wider research community.

Future research should expand this integrative ML framework by incorporating dynamic, time-series physiological measurements and field-based multi-environment trials. Additionally, coupling phenotypic data with transcriptomic and metabolomic profiling will further enhance mechanistic understanding and facilitate gene-to-trait translation [22,23,31]. Such interdisciplinary pipelines could accelerate breeding for resilience, especially under climate-induced drought variability.

With all code and data openly available, this pipeline can now be directly implemented in rice-breeding programs or adapted for high-throughput phenotyping platforms. Ongoing work will focus on piloting these models in field environments and integrating them into breeding pipelines for rapid cultivar screening.

Limitations: While this study demonstrates the predictive value of integrative ML models for drought tolerance, it is constrained by reliance on greenhouse, seedling-stage data and a moderate sample size. Larger, field-based, and multi-environment trials are needed to establish generalizability. Moreover, as the models are based on early phenotypic traits and short-term stress, future research should incorporate time-series and multi-omics data and assess applicability to long-term and reproductive-stage drought responses.

5. Conclusions

This study demonstrated the value of integrating physiological and root anatomical traits with advanced machine-learning models to classify drought tolerance in rice. Through a comparative analysis of random forest, multi-layer perceptron, and a SHAP-optimized stacking ensemble, we achieved high predictive accuracy and identified key biological indicators of drought response. Notably, hydrogen peroxide (H₂O₂), relative water content (RWC), endodermis inner circumference, vascular cylinder thickness, and cortex area consistently emerged as critical features across models.

These findings highlight the importance of oxidative stress signaling and root structural adaptations as central mechanisms of drought resilience in rice. The success of the stacking ensemble approach reinforces the value of ensemble learning in capturing complex, nonlinear trait interactions, while SHAP-based explainability offers new opportunities for interpretable, biology-driven trait selection.

Moving forward, the integration of these insights into breeding pipelines—supported by high-throughput phenotyping and field validation—could accelerate the development of drought-resilient rice cultivars. This work underscores the potential of AI-guided, trait-informed strategies for addressing agricultural challenges under climate change.