基于GIS的输入变量对LSTM模型径流预测的影响

Accurate rainfall-runoff forecasting is crucial in non-urban area,as it is closely related to various human activities,such as grain irrigation,flood warnings and reservoir management.However,the complexity of the terrain and the difficulties in observation make rain-runoff forecasting unfavorable in these areas.To address these challenges,this study employed a Long Short-Term Memory(LSTM)neural network model in conjunction with the data from both Geographic Information System(GIS)and European centre for medium range weather forecasts Re-Analysis version 5(ERA5)to enhance the accuracy of rainfall-runoff prediction.Initially,the new input parameters including Antecedent Soil Moisture Condition(AMC),slope,land use,and soil saturated infiltration coefficient are introduced into the model.The results demonstrate a significant improvement in accuracy through the introduction of the AMC parameter[Nash-Sutcliffe Efficiency Coefficient(NSE)from 0.87 to 0.98].A comparison between the single-input model and the multiple-input ones reveals the latter's superior performance across all parameters.The NSE exceeds 0.9 in three out of four models,indicating strong generalization capabilities.Furthermore,a comparative analysis of multi-parameter models reveals an optimal number of inputs for model performance.It is crucial to consider the size and complexity of study areas when selecting input parameters.