Long Short-Term Memory Neural Network with Attention Mechanism for Rice Yield Early Estimation in Qian Gorlos County, Northeast China

Rice is one of the most extensively cultivated food crops in Northeast China. Estimating pre-harvest rice yield is important for accurately formulating field management strategies and swiftly assessing overall rice production. This can be achieved using a pixel-scale model, which estimates crop yield based on information from each pixel. Previous studies predominantly employed remote sensing indices, climatic data, and yield statistics of rice across either single or all growth periods for yield estimation. These approaches are limited by a delay in yield estimation and are insufficient in the exploration of time-series feature variables at the pixel scale. This study presents the development of a novel deep-learning framework designed for the early estimation of rice yield in Qian Gorlos County, Northeast China. The framework utilizes a long short-term memory neural network integrated with an attention mechanism (ALSTM). In this framework, the heading stage–milk ripening stage is the time window for early yield estimation, and the vegetation indices Normalized Difference Red Edge (NDRE), Green Chlorophyll Vegetation Index (GCVI), and Normalized Difference Water Index (NDWI) from the rice transplanting to the milk ripening stage are time-series feature variables. The yield estimation precision is R 2 = 0.88, RMSE = 341.82 kg/ha, MAE = 280.29 kg/ha, outperforming LASSO (R 2 = 0.71, RMSE = 567.10 kg/ha, MAE = 487.38 kg/ha), RF (R 2 = 0.79, RMSE = 506.70 kg/ha, MAE = 418.90 kg/ha), and LSTM (R 2 = 0.83, RMSE = 451.11 kg/ha, MAE = 326.31 kg/ha). The ALSTM introduced in this paper demonstrates its robustness after being generalized to the 2022 growing season. It can forecast the current-year rice yield two months prior to harvest, providing a valuable reference for developing field management strategies to enhance rice productivity.