Deep learning-based super resolution generative adversarial network for gridded precipitation downscaling across India

Precipitation downscaling is recognized as a challenging problem due to its stochasticity and highly skewed distribution. The downscaling of coarse resolution global climate data is most important in the context of regional-scale climate change impact research and assessments. Image super-resolution is a well-established domain in deep learning, which is also analogous to gridded data downscaling. In this study, the performance of a Super-Resolution Generative Adversarial Network (SRGAN), a cutting-edge deep learning-based image super-resolution technique, is assessed in downscaling and reconstructing high-resolution rainfall data over India from the low-resolution input. The main component of SRGAN is a generator network that takes abstract information from low-resolution (10x10) rainfall data to infer a potential high-resolution (0.250x0.250) counterpart. A super-resolution residual neural network is used as the generator network, which is trained via a supervised learning strategy and adversarial learning strategy. The long-term record of gridded rainfall provided by the India Meteorological Department for the period 1901-2021 is leveraged in this study. A statistical downscaling method called bias correction and spatial disaggregation (BCSD) is also deployed under the same constraints of the train-test split to benchmark deep learning-based downscaling methods. The novel methods are comprehensively assessed for their ability to reconstruct distribution, mean, and extreme rainfall during the test period. Our results show that the deep learning-based methods have an upper hand over the BCSD method for gridded rainfall downscaling over the highly complex terrain of the Indian subcontinent. Keywords: Downscaling, Climate data, Gridded rainfall, Deep learning, SRGAN, Image super-resolution