An interpretable machine learning framework to predict Relative Water Content by using hyperspectral data in Phaseolus vulgaris L. under different irrigation conditions

Common bean (Phaseolus vulgaris, L., 1758) plays a fundamental role in food security, but its high sensitivity to water stress can significantly impact crop productivity. Hence, effective monitoring of plant water status is crucial for optimizing irrigation management. The experiment was conducted in Valenzano (Southern Italy) where bean was grown under three irrigation regimes (full irrigation (FI), deficit irrigation (DI) and rainfed (RD)), arranged in a completely randomized design with 6 replicates. Relative water content (RWC) was measured on three main phenological stages (33, 54 and 63 days after sowing) using the traditional laboratory method that involves weighing leaf samples three times to determine their fresh, saturated and dry weight. Crop spectral reflectance was measured at each phenological stage at three locations for each experimental unit using a handheld hyperspectral radiometer operating in the 325–1075 nm range. Seed yield and water productivity were measured at waxy (74 DAS) and full maturity (120 DAS). The study evaluates the performance of four machine learning (ML) algorithms (Support Vector Regression (SVR), Random Forest (RF), Lasso Regression (LR) and Ridge Regression (RR)) in predicting RWC with two separate data analyses (i) using wavelengths, selected through two feature selection approaches, and then (ii) Vegetation indices (VIs) as predictors. Each algorithm was trained on 80% of the dataset and tested on the remaining 20%, with a grid search cross-validation applied for hyperparameters tuning. When wavelengths were used as predictors, SVR showed the best results (R ± 0.13; RMSE = 6.10 ± 0.55), while RF excelled with VIs (R 2 2 =0.52 = 0.65 ± 0.04; RMSE= 5.37 ± 0.48). SHapley Additive exPlanations (SHAP) values revealed that Red-Edge Vegetation Stress Index (RVSI) and Plant Senes cence Reflectance Index (PSRI) were the most influential VIs in determining RF accuracy. This study could offer a faster and non-destructive approach for assessing water stress and improving irrigation scheduling. The moderate predictive performance of the model (R 2 =0.65) may reflect the non-linear spectral response under water stress, which can also be affected by structural changes, potentially masking water related signals at specific wave lengths; in addition, these results may highlight the need of incorporating SWIR region wavelengths in the models and taking into account the differences between leaf-based indicators and canopy-scale information. These factors provide the basis for further investigation in this research field