Comparative Assessment of Image Super-Resolution Techniques for Spatial Downscaling of Gridded Rainfall Data

With an increasing focus on improving localized understanding of weather and climate phenomena and the computation cost involved in high-resolution modelling, spatial downscaling of data has proven to be a viable alternative method to obtain high-resolution climate data. Historically, several statistical and dynamical downscaling techniques have been employed to enhance coarse-resolution data to a finer grid scale. However, these conventional methods are either inefficient in capturing several finer scale details or are resource-heavy for recursive applications. Image Super-Resolution (SR) is a computer vision concept of using grid-based approaches to enhance the resolution of an image, analogous to spatial downscaling. In this study, we explore and compare the performance of different image super-resolution methods in downscaling the Indian Meteorological Department’s (IMD) gridded rainfall data from a low spatial resolution (1.0°) to a high resolution (0.25°). The SR methods considered include a fully connected autoencoder, four traditional convolutional neural networks, and two residual neural networks. The primary objective of this study was to make an initial assessment of the performance of such methods in downscaling gridded rainfall data over the Indian region. Furthermore, the resultant downscaled products have been compared using different objective metrics and analytical comparisons with ground truth data. One of the key findings of this study is that the residual learning-based neural networks demonstrated better performance in creating perceptually realistic rainfall maps (proven both quantitatively and qualitatively), closely followed by the traditional convolutional neural networks and the autoencoder.