Xin Tian, Fulang Cen, Junshuo Wei, Quanzhi Zhao, Zhenran Gao
Rice is a major staple crop that feeds nearly half of the global population; thus, its accurate yield estimation is critical for food security and agricultural management. To improve prediction accuracy and applicability, in this study, we propose a multi-temporal fusion framework that integrates unmanned aerial vehicle (UAV)-based multispectral and thermal infrared data with feature selection and deep learning models. UAV data were collected from 57 rice varieties at four key growth stages (tillering, heading, grain filling, and maturity). Vegetation indices (VIs), texture feature indices (TFIs), and canopy temperature (T) were extracted, and a Random Forest (RF) model was employed to identify dominant features. Subsequently, temporal feature sequences were constructed to develop yield prediction models under single-, dual-, and multi-temporal scenarios. Five deep learning models, along with an RF baseline, were evaluated. This study demonstrates a feature-driven deep learning framework effective for small-to-moderate datasets. Our results indicate that the convolutional neural network–long short-term memory (CNN–LSTM) model achieved the best performance, with a coefficient of determination (R 2 ) of 0.87 and root mean square error (RMSE) of 0.55 kg/m 2 under multi-temporal conditions. The integration of VIs, TFIs, and T consistently outperformed single-source features, thereby highlighting the advantages of combining spectral, structural, and thermal information. Among individual growth stages, the heading stage yielded the best single-temporal performance (R 2 = 0.83), whereas the combination of tillering, heading, and grain filling achieved the highest overall accuracy. In this study, we provide an effective framework for UAV-based rice yield prediction and offer practical guidance for selecting optimal observation stages and data combinations, thereby supporting precision agriculture and crop management.