MAUNet: a max-average neural network architecture for precipitation downscaling

Published: 01 Jan 2024, Last Modified: 06 Feb 2025Neural Comput. Appl. 2024EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Most operational weather and climate models carry out forecasts or simulations of atmospheric variables at low spatial resolutions, but we often need high-resolution projections of such variables. The process of mapping low-resolution projections to high-resolution projections is called spatial downscaling. This is analogous to the computer vision task of single-image super-resolution (SISR). In recent studies, convolution-based architectures including UNet and its variants have emerged as a good choice for SISR. Since the gridded spatial map of any climate variable is analogous to an image, we can use the SISR-based models for spatial downscaling. In this paper, we present a novel UNet-based architecture called max-average UNet (MAUNet) to downscale the precipitation. We have proposed Max-Average Units (MAUs) that include a max-pooling, an average-pooling, and an averaging unit for the encoding part of the UNet. For the decoding part, we develop UpSampler Unit (USU), which too utilizes averaging. We demonstrate the importance of max-averaging units through toy experiments on the MNIST and CelebA dataset. We also examine the role of dropouts for this task and experimentally demonstrate their skill to improve the model’s performance on different datasets. The task examined here is to obtain fourfold resolution enhancement of monsoon precipitation over the Indian landmass and also over Southeast Continental USA (CONUS) region. The models are evaluated using RMSE, PSNR, MSSIM, and correlation coefficient as evaluation matrices. We have made detailed comparisons to show that MAUNet can produce superior downscaling than standard interpolation techniques as well as previously used deep learning models like UNet, EDSR, and SRDRN.
Loading