Multitask ST-LSTM model based on UAV hyperspectral remote sensing for wheat yield prediction

Timely and accurate acquisition of wheat yield is essential for ensuring food security and improving breeding efficiency. However, temporal variation relationships across growth stages have not been fully exploited in existing wheat yield prediction methods. Moreover, prediction accuracy and stability are easily affected by environmental factors and distribution shifts across different yield levels. To address these limitations, this study proposed a multitask stage temporal long short-term memory (ST-LSTM) model. A spectral- and temporal-level data augmentation method was first applied in the proposed framework to preserve the characteristics of wheat growth stages while expanding the limited sample size. Furthermore, wheat growth trend information was incorporated into the feature selection process, with both local hyperspectral features and temporal dependencies across growth stages taken into account. Subsequently, a multitask architecture combining variety classification and yield prediction was constructed, and a yield constraint was introduced to enhance the consistency between predicted yield and yield-level structure. In this framework, the yield-level probabilities learned from the variety classification task, stage-temporal features, and prior knowledge of yield distribution were jointly used to optimize the yield prediction task. Experiments on two-year datasets and an independent third-year dataset showed that the proposed model outperformed single-task models and conventional temporal models in both variety classification and yield prediction, while maintaining good stability in cross-year prediction.