Comparison between 2D-CNN a 3D-CNN for precipitation downscaling

Abstract:

Future climate projections are important for disaster prevention under a changing climate. However, the resolutions of most future climate projections data are too coarse, and then downscaling is required for regional-scale analysis. There are two major downscaling techniques: dynamical downscaling and statistical downscaling. In addition to them, a deep learning method is nowadays utilized for downscaling. Two-Dimensional Convolutional Neural Network (2D-CNN) is a deep learning method that was successfully applied to precipitation downscaling. In the application atmospheric reanalysis data were used as the inputs for precipitation downscaling. Atmospheric processes have a three-dimensional structure. In this study, therefore, atmospheric reanalysis data were also utilized as input and the basin-scale precipitation was used as the target data. Then, this study employed the three-dimensional CNN (3D-CNN) for precipitation downscaling, and compared it with 2D-CNN. 3D-CNN may be able to extract the three-dimensional features of the input data. For the comparison, both of 2D-CNN and 3D-CNN were implemented at a study area for precipitation downscaling. As a study area, the Shira-River Basin, which is located in the Kyushu region, Japan, was selected. The basin average values of precipitation were used as the target variable of the precipitation downscaling. The results showed that 3D-CNN has potential to improve the accuracy of precipitation downscaling, compared to 2D-CNN.