Improving Accuracy and Interpretability of Rice Aboveground Biomass Estimation by Integrating <scp>CatBoost</scp> and <scp>SHAP</scp> Model

ABSTRACT Accurate and efficient assessment of rice aboveground biomass (AGB) is crucial for enhancing field management, yield prediction, and decision‐making in precision agriculture. While machine learning (ML) techniques have significantly improved the efficiency of AGB estimation, their inherent “black‐box” nature often restricts model interpretability. This study aims to enhance both the interpretability and accuracy of rice AGB estimation models by integrating the CatBoost ensemble learning algorithm with SHapley Additive exPlanations (SHAP). An unmanned aerial vehicle (UAV) platform was used to capture canopy multispectral images of rice throughout its entire growth cycle under diverse field conditions. A high‐precision CatBoost model for rice AGB estimation was developed based on extracted vegetation indices (VIs) and texture features (TFs). SHAP analysis was applied to quantitatively assess the impact of input features and their interactions on AGB estimation. The results demonstrated that the CatBoost model, incorporating VIs and TFs, outperformed both the random forest regression (RFR) and LightGBM models, achieving optimal predictive performance when trained on 90% of the dataset ( R 2 = 0.96, RMSE = 813.00 kg/ha). SHAP analysis revealed that TFs (mean, homogeneity, variance, and correlation) and VIs (visible atmospherically resistant rededge indices (VARIre), normalized difference rededge index (NDRE) and normalized difference vegetation index (NDVI)) were the primary factors influencing AGB estimation. The main and interaction effects of input features contributed 76% and 44% to AGB estimation on the testing set, and 59% and 55% on the training set, respectively. This study offers a reliable and cost‐effective method for AGB estimation and provides an interpretable predictive framework for broader agricultural remote sensing applications.