Climate change and water scarcity bring significant challenges to agricultural systems in the Mediterranean region. Novel methods are required to rapidly monitor the water stress of the crop to avoid qualitative losses of agricultural products. This study aimed to predict the stem water potential of cotton (Gossypium hirsutum L., 1763) using Sentinel-2 satellite imagery and machine learning techniques to enhance monitoring and management of cotton's water status. The research was conducted in Rutigliano, Southern Italy, during the 2023 cotton growing season. Different machine learning algorithms, including random forest, support vector regression, and extreme gradient boosting, were evaluated using Sentinel-2 spectral bands as predictors. The models' performance was assessed using R2 and root mean square error (RMSE). Feature importance was analyzed using permutation importance and SHAP methods. The random forest model using Sentinel-2 bands' reflectance as predictors showed the highest performance, with an R2 of 0.75 (±0.07) and an RMSE of 0.11 (±0.02). XGBoost (R2: 0.73 ± 0.09, RMSE: 0.12 ± 0.02) and AdaBoost (R2: 0.67 ± 0.08, RMSE: 0.13 ± 0.02) followed in performance. Visible (blue and red) and red edge bands were identified as the most influential predictors. The trained RF model was used to model the seasonal trend of cotton's stem water potential, detecting periods of acute and moderate water stress. This approach demonstrates the prospective for high-frequency, non-invasive monitoring of cotton's water status, which could support smart irrigation strategies and improve water use efficiency in Mediterranean cotton production.
Keywords: Gossypium; Optuna; drought stress; machine learning; remote sensing; satellite.