Comparative analysis of deep learning architectures trained for downscaling gridded precipitation across India

For understanding localized hydrological and climatological processes, downscaling gridded precipitation data to finer spatial resolutions is a crucial prerequisite. For a densely populated country like India, accurate downscaled data is crucial for building resilience to climate change impacts, supporting adaptation efforts, and enhancing disaster management. In recent years, deep learning (DL) has emerged as a powerful tool for advancing Earth system modelling and climate data downscaling. This study presents a comprehensive intercomparison of deep learning architectures, for downscaling precipitation across India. A few efficient DL architectures from recent studies are chosen for intercomparison such as simple dense, simple convolutional neural network, Fast Super Resolution Convolutional Neural Network (FSRCNN), Super Resolution Deep Residual Network (SRDRN), U-Net, and Nest-U-Net. The experiments are designed in synthetic style by using coarsened ECMWF Reanalysis version 5 (ERA5; 1^ox1^o) daily variables as the inputs and high-resolution Indian Monsoon Data Assimilation and Analysis reanalysis (IMDAA; 0.12^ox0.12^o) daily precipitation as training labels and benchmarks for the evaluation. Training and validation are conducted for the period 1980-2014, afterwards the trained models are evaluated on data from 2015-2020. To reduce the biases induced by the highly positive-skewed precipitation data and to enhance the model performance on extreme events, a weighted mean absolute error is implemented for training. The performance of the DL models is also compared with the Bias Correction and Spatial Disaggregation (BCSD), a renowned statistical downscaling method. The results indicate that all deep learning DL models outperformed the BCSD method. Among the DL models, U-Net and Nest-U-Net demonstrated superior performance in capturing fine-scale precipitation patterns and extreme precipitation events, owing to their encoder-decoder architecture, which effectively learns spatial features at different scales. In contrast, the FSRCNN and SRDRN produced results with slightly lower precision than the U-Net models, but at a significantly reduced inference time, making them more efficient for faster data generation. The findings underscore the potential of deep learning for improving regional precipitation downscaling across India, offering a promising alternative to traditional statistical methods like BCSD in handling complex, non-linear relationships inherent in climate data.