Downscaling and reconstruction of high-resolution gridded rainfall data over India using deep learning-based generative adversarial network

Abstract

To expedite regional-scale climate change impact research and assessments, the downscaling of climate data is a crucial prerequisite. Image super-resolution, which is analogous to gridded data downscaling, is the concept of improving the pixel quality of images using deep learning techniques. 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 producing perceptually realistic 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 (LR) rainfall data to infer potential high-resolution (HR) counterparts. A Super-Resolution Residual Neural Network (SRResNet) is used as the generator network. It is trained using a supervised learning strategy (SRResNet) and adversarial learning strategy (several variants created, e.g., SRGAN-MSE, SRGAN-VGGB2, SRGAN-VGGB3 and SRGAN-VGGB4). A statistical downscaling method called bias correction and spatial disaggregation (BCSD) is also employed to compare with the deep learning-based downscaling methods. All these methods are rigorously assessed for their ability to reconstruct distribution, mean, and extreme rainfall during the test period. Our results show that the supervised learning-based SRResNet and adversarial learning-based SRGAN-MSE variant has an upper hand over the BCSD method for gridded rainfall downscaling. These findings have important implications for enhancing the precision and quality of regional climate data in the context of climate change impact assessment.

More information

Identifiers

ORCID for Sreevathsa Golla: orcid.org/0000-0003-4084-9677

Catalogue record

Export record

Altmetrics