DownScaleBench for developing and applying a deep learning based urban climate downscaling- first results for high-resolution urban precipitation climatology over Austin, Texas

Components of the DownScaleBench framework for generating supervised learning datasets for urban downscaling. An integral part of DownScaleBench is the incorporation of station information in the training process

Abstract

This study introduces DownScaleBench, a novel deep learning approach for urban downscaling, providing detailed climate data essential for resilient infrastructure development and adaptation planning. Using ground-based and satellite/reanalysis data, it employs an iterative super-resolution convolutional neural network (Iterative SRCNN) to enhance resolution, producing high-resolution (300 m) precipitation datasets from coarse (10 km) satellite-based products like JAXA GsMAP. Validation against in-situ observations during heavy rain events in Austin, Texas, demonstrates significant improvement over baseline cubic interpolation methods. DownScaleBench offers promise for creating high-resolution urban meteorological datasets, crucial for climate-resilient city planning efforts.


Team

Manmeet Singh, Nachiketa Acharya, Sajad Jamshidi, Junfeng Jiao, Zong‑Liang Yang, Marc Coudert, Zach Baumer and Dev Niyogi

Replication data and code can be found at https://github.com/texuslabut/ urban precipitation downscaling.

Code availability

Funding

This work was beneftted from the National Aeronautics and Space Administration (NASA) Interdisciplinary Science (IDS) 80NSSC20K1262 and 80NSSC20K1268, the National Science Foundation (NSF) Grants for Rapid Response Research (RAPID) AGS 19046442, and the Department of Energy (DOE) Advanced Scientifc Computing Research Program Grant No. DE-SC002221.

The cover image is sourced from Pexels, is free of copyright issues, and can be used for educational purposes. https://help.pexels.com/hc/en-us/articles/360042295174