Study on Large-Scale Crop Disease Epidemic Forecasting Based on a Fusion Model Between SEIR and LSTM Networks

Establishing effective models for dynamically forecasting crop diseases is crucial for supporting grain production in China. Addressing the challenge of fusing the mechanistic model and multi-sources data, this paper proposes the SEIR-LSTM model, which integrates the SEIR mechanistic model with Long Short-Term Memory (LSTM) network. The SEIR model captures disease trends based on biological mechanisms, while LSTM models nonlinearly represent multiple driving factors, achieving high-precision dynamic disease forecasting. The model is validated across counties in Hunan Province, China. Experimental results demonstrate that the integrated SEIR-LSTM model effectively captures disease trends and nonlinear patterns between driving factors and disease epidemic. In cross-county forecasting, the average RMSE is 4.716, the average MAE is 4.116, and the average r is 0.706, indicating high forecasting accuracy and regional generalization. This model can provide effective decision-making support for crop management practices.