Rice yield prediction base on UAV multispectral imagery using machine learning methods

• Multi-temporal drone data from crucial rice growth stages can significantly improve the accuracy of rice yield prediction. • The integration of meteorological data enhances yield prediction accuracy by 10.7-12.9% compared to the sole use of vegetation indices. • SHAP analysis identifies heading and tillering stages as critical for yield determination. • The XGB model demonstrates strong generalizability across diverse agroecological zones in China. Accurate rice yield prediction is critical for sustainable agriculture and food security, yet traditional models often fail to capture complex environmental interactions. This study aims to develop a field-scale rice yield prediction model integrating multi-temporal unmanned aerial vehicles (UAV) multispectral data and meteorological data across three agroecological zones in China from 2022 to 2024. Using UAV multispectral data, 69 vegetation indices (VIs) were constructed and evaluated for three rice cultivars (Ningxiangjing9, Taian1, and Yinxiang38) under varying nitrogen treatments, and the optimal VIs were selected for five rice growth stages. Ten meteorological features (MF) per growth stage, including temperature and precipitation, were integrated into rice yield prediction models to account for environmental influences. In evaluating the prediction of rice yield, six regression models were assessed: simple linear regression (SLR), partial least squares regression (PLSR), random forest (RF), gaussian process regression (GPR), support vector machine (SVM), and extreme gradient boosting (XGB). Among models, the five-period XGB model, integrating multi-temporal vegetation indices (VIs) and MF, achieved the highest accuracy (R 2 = 0.83). SHapley Additive exPlanations (SHAP) analysis identified normalized difference red edge (NDRE) at heading stage and triangular chlorophyll index/optimized soil-adjusted vegetation index (TCI/OSAVI) at tillering stage as the most influential features, with precipitation and temperature also significant. The integration of MF improved accuracy by 10.7-12.9% relative to models that relied solely on VIs. These results highlight the potential of using multi-temporal VIs and MF with machine learning algorithms for precise rice yield prediction, supporting precision agriculture across diverse agroecological conditions.