Downscaling of GCM Output Using Deep Learning Techniques: Fid-Bau Portal

Downscaling of GCM Output Using Deep Learning Techniques

Tamang, Chandra Prakash / Paul, Subir / Kumar, Dasika Nagesh

General circulation models (GCMs) provide the data to study climate change under different scenarios, but they operate on a coarse scale. Therefore, to assess the hydrological impacts of global climate change on a regional scale, the output from a GCM must be downscaled to finer resolution. Moreover, regional precipitation simulations can be improved by using physically relevant variables from GCMs at different pressure levels as predictors. In this study, we have explored different deep learning techniques for multi-site downscaling of daily precipitation over the Mahanadi basin using large-scale hydroclimate variables as predictors. Hydroclimatic variables from NCEP reanalysis data (available at 2.5° × 2.5° resolution) are used to downscale the daily precipitation product at observational grid-scale (i.e., 1° × 1° spatial resolution). Different deep learning architectures, viz., deep neural network (DNN), 2D- and 3D-convolutional neural network (CNN), and hybrid-DNN are trained on these spatio-temporal variables. The results show that deep learning models have the ability to use the spatial information from predictor variables over the Indian subcontinent to capture monsoon patterns and downscale daily precipitation. The 2D-CNN model is able to learn the spatial features from high-dimensional predictor variables over continental sized domains. 3D-CNN further reduces the number of parameters and is able to learn from the stacked high-dimensional spatio-temporal datasets at different vertical pressure levels with comparative ease. The hybrid-DNN is employed to make use of spatial structure of the predictor datasets as well as the information from the GCM precipitation outputs of neighboring grid points of the observation grids. The 2D-CNN, 3D-CNN, and hybrid-DNN perform better than the DNN showing the usefulness of exploiting the spatial gridded structure of the predictors. This study highlights the potential of deep learning techniques in learning precipitation patterns from coarse-resolution climate model outputs and in downscaling daily precipitation.