Abstract
Accurate precipitation monitoring is critical for disaster mitigation, yet global satellite products often suffer from significant bias and low spatial resolution. We present PrecipDiff, a novel framework that utilizes Denoising Diffusion Probabilistic Models (DDPM) and residual learning to simultaneously correct biases and downscale satellite observations from 10km to 1km resolution. By focusing on learning residuals rather than direct precipitation values, PrecipDiff achieves state-of-the-art accuracy and visual fidelity using only precipitation data, offering a cost-effective alternative to ground-based radar in data-scarce regions.
Introduction
Water-related disasters, including floods and landslides, account for the vast majority of natural disaster fatalities, particularly in developing nations. While ground-based Weather Surveillance Radar (WSR) provides high-resolution, accurate data, its high cost limits its deployment globally. Satellite-based Precipitation Products (SPPs) like IMERG provide global coverage but are hindered by systematic biases and a coarse spatial resolution (typically 10km), which is insufficient for localized urban planning and flood warnings. PrecipDiff addresses these challenges by treating precipitation enhancement as a computer vision task, leveraging generative diffusion models to "sharpen" and "calibrate" satellite data.
Methodology
PrecipDiff employs a two-stage pipeline based on Denoising Diffusion Probabilistic Models (DDPM). Unlike traditional regression models, PrecipDiff utilizes a residual learning approach, which focuses on generating the difference (residual) between low-quality and high-quality data.
The framework consists of:
- Bias Correction: The model learns to generate the residual between biased satellite observations (IMERG) and coarsened ground-truth radar data (MRMS). This step aligns the statistical distribution of satellite data with ground truth.
- Downscaling: The corrected 10km data is upsampled, and a second diffusion model generates high-frequency residuals to restore fine-scale (1km) details.
The architecture utilizes a U-Net for denoising, conditioned on low-resolution inputs, and the EDM scheduler for efficient inference. A key advantage of this approach is its reliance solely on precipitation data, eliminating the need for auxiliary meteorological variables like elevation or temperature.
Datasets and Experimental Setup
The models were trained using data from the Seattle region (2022-2023), utilizing IMERG as the low-resolution satellite input and MRMS as the high-resolution ground truth. To test generalizability, the framework was evaluated on "Out-of-Distribution" data from New York, Portland, and San Jose, representing diverse climatic conditions.
Results and Discussion
PrecipDiff significantly outperforms standard interpolation and transformer-based super-resolution baselines. In the unified task of correcting and downscaling raw 10km satellite data to 1km, PrecipDiff achieved an RMSE of 1.0846, a substantial improvement over the raw satellite data's RMSE of 3.3410.
As shown in the error distribution analysis, the model effectively mitigates the common satellite bias of overestimating light rain and underestimating heavy precipitation. Qualitatively, the diffusion process recovers realistic storm structures and sharp fronts that are otherwise lost in coarse satellite pixels.
Conclusion and Future Work
PrecipDiff demonstrates the power of diffusion models in environmental monitoring. While the downscaling component generalizes well across regions, the bias correction module shows higher sensitivity to local geography, suggesting a need for more diverse training data. Future work will focus on developing a global foundation model for precipitation to address these regional distribution shifts and improve intensity estimation for extreme weather events.
Summary updated