GeoRain v2.0: Enhancing Radar Rainfall Maps Using a Diffusion-Based Generative AI Approach with Satellite Imagery

In the era of the climate crisis, timely and accurate rain forecasting has become essential for life-saving early warning systems. With extreme rainfall events on the rise, the risk of secondary disasters like severe flooding and landslides is escalating. Traditional weather radar systems, although effective in monitoring unexpected heavy rainfall, are expensive to install and maintain. Moreover, their coverage is often limited to land areas. To address these limitations, we developed GeoRain, an AI-based, real-time rainfall measurement products that leverages data from geostationary satellites. This product is tailored for rainfall monitoring in areas that lack access to conventional radar systems.

Our study validates that GeoRain can effectively serve as a substitute for ground-based radar systems, consistently outperforming existing satellite-based rainfall measurement products. Initially, GeoRain was built on a Generative Adversarial Network (GAN) framework, specifically employing the Pix2PixCC model. We have since refined GeoRain by integrating a state-of-the-art diffusion-based image generative model to derive proxy radar images from satellite data.

The first iteration, GeoRain v1.0, translates imagery from the geostationary GK2A satellite into radar reflectivity, generating single images based on input datasets. In contrast, the more advanced GeoRain v2.0 employs diffusion-based methods to produce a broader range of sample outputs. These generated datasets can be systematically ensembled to offer robust predictions for rare yet extreme rainfall events. Notably, GeoRain v2.0 demonstrates a substantial reduction in Mean Squared Error (MSE) in measuring rainfall intensity compared to its predecessor, GeoRain v1.0.

This advanced model provides a fast solution for generating radar-like rainfall data in areas lacking ground-based radar systems, extending its utility to maritime regions as well. The substantial improvements in performance metrics highlight GeoRain as a significant contribution to the field of rainfall prediction and environmental monitoring, particularly in a dynamically changing climate.