Combining UAV-Based Multispectral and Thermal Infrared Data with Regression Modeling and SHAP Analysis for Predicting Stomatal Conductance in Almond Orchards

Abstract

Understanding and accurately predicting stomatal conductance in almond orchards is critical for effective water-management strategies, especially under challenging climatic conditions. In this study, machine-learning (ML) regression models trained on multispectral (MSP) and thermal infrared (TIR) data acquired from unmanned aerial vehicles (UAVs) are used to address this challenge. Through an analysis of spectral indices calculated from UAV-based data and feature-selection methods, this study investigates the predictive performance of three ML models (extra trees, ET; stochastic gradient descent, SGD; and extreme gradient boosting, XGBoost) in predicting stomatal conductance. The results show that the XGBoost model trained with both MSP and TIR data had the best performance (R2 = 0.87) and highlight the importance of integrating surface-temperature information in addition to other spectral indices to improve prediction accuracy, up to 11% more when compared to the use of only MSP data. Key features, such as the green–red vegetation index, chlorophyll red-edge index, and the ratio between canopy temperature and air temperature (Tc-Ta), prove to be relevant features for model performance and highlight their importance for the assessment of water stress dynamics. Furthermore, the implementation of Shapley additive explanations (SHAP) values facilitates the interpretation of model decisions and provides valuable insights into the contributions of the features. This study contributes to the advancement of precision agriculture by providing a novel approach for stomatal conductance prediction in almond orchards, supporting efforts towards sustainable water management in changing environmental conditions.