PrecipDiff: Leveraging image diffusion models to enhance satellite-based precipitation observations

TL;DR

The paper tackles the scarcity of ground-based precipitation monitoring by enhancing satellite products through a diffusion-based residual learning framework called PrecipDiff. It formulates bias correction and downscaling as two residual-learning tasks, training two diffusion models with precipitation-only data and an EDM scheduler, to produce high-resolution outputs from 10 km IMERG data toward 1 km MRMS-like detail. In Seattle (2022–2023), the approach outperforms a state-of-the-art baseline in RMSE and CRPS for both correction and downscaling, and the unified pipeline achieves the best numerical alignment with radar data while delivering richer fine-scale structure. The method holds promise for operational use in data-scarce regions, enabling near-real-time, high-resolution precipitation estimates without auxiliary variables.

Abstract

A recent report from the World Meteorological Organization (WMO) highlights that water-related disasters have caused the highest human losses among natural disasters over the past 50 years, with over 91\% of deaths occurring in low-income countries. This disparity is largely due to the lack of adequate ground monitoring stations, such as weather surveillance radars (WSR), which are expensive to install. For example, while the US and Europe combined possess over 600 WSRs, Africa, despite having almost one and half times their landmass, has fewer than 40. To address this issue, satellite-based observations offer a global, near-real-time monitoring solution. However, they face several challenges like accuracy, bias, and low spatial resolution. This study leverages the power of diffusion models and residual learning to address these limitations in a unified framework. We introduce the first diffusion model for correcting the inconsistency between different precipitation products. Our method demonstrates the effectiveness in downscaling satellite precipitation estimates from 10 km to 1 km resolution. Extensive experiments conducted in the Seattle region demonstrate significant improvements in accuracy, bias reduction, and spatial detail. Importantly, our approach achieves these results using only precipitation data, showcasing the potential of a purely computer vision-based approach for enhancing satellite precipitation products and paving the way for further advancements in this domain.

Figures (8)

• Figure 1: Overview flowchart for inference process. Satellite precipitation data are correctized by the first diffusion process and then upsampled using a linear method. Second diffusion process are applied to synthesized downscaled predictions to create high-resolution data.
• Figure 2: Overview flowchart for training correction and downscaling diffusion models. The correction model is trained at 10 km resolution using noisy residuals as input and IMERG data as conditioning information. Residuals at this stage are calculated by subtracting IMERG data from LR MRMS data. The downscaling model is trained at 1 km resolution, with residuals computed by subtracting interpolated LR MRMS data from original MRMS data.
• Figure 3: A comparison of the error distribution between original IMERG and Corrected IMERG. The errors are calculated based on the LR MRMS observations, with all pixels flattened for the comparison.
• Figure 4: A comparison of the error distributions between LR MRMS and downscaled MRMS. The errors are calculated based on the original MRMS observations, with sampling from the pixels flattened for the comparison.
• Figure 5: Illustration of a unified test for enhancing the satellite precipitations. The IMERG is calibrated by the corrector resulting in (b) corrected IMERG, and (c) is sampled by conditioning on the corrected IMERG. A comparison is made between (c) corrected and downscaled IMERG and (d) MRMS data.