An Interpretable Deep Learning Model For Surface Hydrological Processes At Multiple Time Scales

Deep learning has been successfully applied to hydrology and has made significant progress in the prediction of nonlinear and non-stationary hydrological processes.Although some studies use machine learning methods in streamflow forecasting,few studies systematically compare the forecasting performance of models at different time scales and gain insight into the impact of model storage mechanisms on forecasting performance.In the streamflow forecasting work,a long short-term memory RNN LSTM model composed of units,input gates,output gates and forgetting gates is built,which can learn long-term time dependencies and capture nonlinear relationships for forecasting daily,weekly and monthly time scale of streamflow.Deep learning models can be trained and validated using multiple datasets of climatic and hydrological elements from the Daomaguan Hydrological Station in the Tang River Basin from 2006 to 2014.By setting multiple timing steps,the effect of the length of memory cell storage information on different time scales is explored.We use the RNN LSTM model to simulate and forecast the surface hydrological processes in the Tang River Basin on three specified time scales: daily,weekly and monthly.The results show that the RNN LSTM model has good performance in daily scale prediction,and the coarse-grained characteristics of the dataset are the key factors affecting the weekly and monthly scale prediction performance.In order to make the deep learning model explicable,this study established three explicable deep neural network models.We used Shapley Addictive ex Planation derived from game theory to explain the output of the model and use SHAP to analyze the contribution of each climatic factor at any given time output at the specified future time based on Shapley values.Shapley allows one to rank the contributions of the climatic factors to streamflow,to identify active times of the climatic factors and to quantify contributions of the climatic factors to the peak and low flow.The coupling of the LSTM model with SHAP creates an interpretable deep neural network model for analyzing surface geological processes in the Tang river basin,acquiring a great deal of insights into surface hydrological processes in the studied area.It can be readily generalized to other geological locations and processes as well.