Deep Learning Based Statistical Downscaling for Enhanced Representation of Indian Monsoon Rainfall With Subseasonal Variability and Extremes

Abstract

Indian summer monsoon rainfall (ISMR) accounts for 80% of India's annual rainfall and impacts 1.4 billion people, yet remains poorly simulated by climate models. Although super resolution deep learning models such as YNet show promise in global downscaling efforts, we found their performance on ISMR less effective, particularly for extreme rainfall and intraseasonal variability. The transferability of downscaling algorithms trained on reanalysis data to climate model simulations, as well as a comprehensive evaluation across key timescales of ISMR variability, remain unexplored. To address these limitations, we propose two variants of YNet: YNet_D, which incorporates atmospheric variables to enhance downscaling accuracy, and YNet_DE, which prioritizes extreme rainfall using a weighted loss function. YNet_D and YNet_DE outperform YNet, reducing root mean squared error from 12.41 to 12.25 mm and 11.97 mm, respectively. For extremes, they lower the R95p bias from 78.27 to 56.52 mm and 40.22 mm. Both models also show improved performance at intraseasonal and interannual timescales and demonstrate better transferability to General Circulation Models. Unlike conventional statistical approaches, the models proposed retain key physical dynamics, providing a robust solution that preserves the system's variability and complexity.