CHARACTERISTICS OF PRECIPITATION DOWNSCALING BY MEANS OF DEEP LEARNING METHOD

Abstract

 The objective of this study is to investigate characteristics of precipitation downscaling by means of a deep learning method, which have not been revealed well so far. This study utilized atmospheric variabes as input, and watershed-scale precipitation as the targeted data. CNN was seleted as a deep learing method for precipitation downscaling. Then, this study analyzed impacts of the spatial range obtaining input variables and the selection of combination of input variables on the accuracy of estimated precipitation. The results indicates several characteristics of CNN precipitation downscaling (CNN-DS) : (1) A wide horizontal range of the input variables are required compared to the target area. (2) CNN-DS requires physical information to some extent. (3) Depending on a variable, it may decrease the accuracy. (4) The use of more input variables related to precipitation does not always improves the accuracy.