An Intercomparison of Deep-Learning Methods for Super-Resolution Bias-Correction (SRBC) of Indian Summer Monsoon Rainfall (ISMR) Using CORDEX-SA Simulations

Abstract

The Indian Summer Monsoon Rainfall (ISMR) plays a significant role in India’s agriculture and economy. Our understanding of the climate dynamics of the Indian summer monsoon has been enriched with general circulation models (GCMs) and regional climate models (RCMs). Systematic bias associated with these numerical simulations, however, needs to be corrected before we can obtain accurate or reliable projections of the future. Therefore, this study applies two state-of-the-art deep-learning (DL)-based super-resolution bias correction (SRBC) methods, viz. Autoencoder-Decoder (ACDC) and a deeper network Residual Neural Network (ResNet) to perform spatial downscaling and bias-correction on high-resolution CORDEX-SA climatic simulations of precipitation. To do so, we obtained eight meteorological variables from CORDEX-SA RCM simulations along with a digital elevation model at a spatial resolution of 0.25°×0.25° as input. Indian Monsoon Data Assimilation and Analysis, precipitation reanalysis re-grided to 0.05°×0.05° spatial resolution is chosen as output for the training period 1979–2005. To evaluate the DL algorithms, the RCP 2.6 scenario of CORDEX-SA future simulations for the period 2006–2020 is chosen. Moreover, we also conducted a performance assessment of the representation of mean, variability, extreme, and frequency of rainfall associated with ISMR. The results of the experiments show that the DL method ResNet a highly efficient in (i) improving the spatial resolution of the climatic simulations from 0.25°×0.25° to 0.05°×0.05°, (ii) reducing the systematic biases of the extreme rainfall of ISMR from 21.18 mm to -7.86 mm, and (iii) providing a robust bias-corrected climate simulation of ISMR for future climate mitigation and adaptation studies.