2021 Volume 2 Issue J2 Pages 883-892
This study conducted sensitivity analysis of Long Short-term memory (LSTM) network to its input var-iables. Rainfall-runoff modeling at a snow-dominated watershed was targeted as a case study. As the sen-sitivity analysis, this study investigated the effect of each input variable such as air temperature, radiation, wind, surface pressure, and evapotranspiration on the accuracy of the simulated flow discharge by LSTM. Meanwhile, this study conducted discussions on the effects on the accuracy in consideration with the phys-ical relations among the input variables and flow discharge. The results showed that input variables related to hydrological processes do not always improve the model accuracy. In recent years, there has been a demand for interpretation of learning results by deep learning methods from a physical perspective. Without revealing the relations between the input variable and the estimation accuracy, such interpretation may not be reliable.
