On the trendy deep studying approaches for precipitation downscaling
Authors: Bipin Kumar, Kaustubh Atey, Bhupendra Bahadur Singh, Rajib Chattopadhyay, Nachiket Acharya, Manmeet Singh, Ravi S. Nanjundiah, Suryachandra A. Rao
Summary: Deep Studying (DL) based mostly downscaling has grow to be a well-liked instrument in earth sciences lately. More and more, totally different DL approaches are being adopted to downscale coarser precipitation information and generate extra correct and dependable estimates at native (~few km and even smaller) scales. Regardless of a number of research adopting dynamical or statistical downscaling of precipitation, the accuracy is restricted by the supply of floor fact. A key problem to gauge the accuracy of such strategies is to check the downscaled information to point-scale observations which are sometimes unavailable at such small scales. On this work, we stock out the DL-based downscaling to estimate the native precipitation information from the India Meteorological Division (IMD), which was created by approximating the worth from station location to a grid level. To check the efficacy of various DL approaches, we apply 4 totally different strategies of downscaling and consider their efficiency. The thought-about approaches are (i) Deep Statistical Downscaling (DeepSD), augmented Convolutional Lengthy Quick Time period Reminiscence (ConvLSTM), absolutely convolutional community (U-NET), and Tremendous-Decision Generative Adversarial Community (SR-GAN). A customized VGG community, used within the SR-GAN, is developed on this work utilizing precipitation information. The outcomes point out that SR-GAN is the very best methodology for precipitation information downscaling. The downscaled information is validated with precipitation values at IMD station. This DL methodology affords a promising various to statistical downscaling.
