Tag: Global Satellite Mapping of Precipitation (GSMaP)

Researchers at Chiba University (led by Assistant Professor Daiko Kishikawa, with Dr. Yuka Muto and Dr. Shunji Kotsuki) have developed a machine- learning method to fill gaps in global satellite- based rainfall maps.

Global rainfall maps, such as those from the Japan Aerospace Exploration Agency’s Global Satellite Mapping of Precipitation (GSMaP), are assembled from multiple satellites. Because satellites cannot observe the entire Earth at once, some areas lack valid observations. Traditional gap- filling approaches rely on earlier and later observations but often fail to fully capture how precipitation patterns evolve in space and time, leading to inconsistencies.

The team reframed the problem as a video- reconstruction (inpainting) task. They treat sequences of precipitation maps like video frames and estimate missing sections using information from nearby times and surrounding regions, plus auxiliary data. Their approach combines a three- dimensional (3D) U- Net convolutional neural network (which captures spatiotemporal patterns) with a 3D condition encoder. The encoder incorporates infrared satellite images, geographic coordinates, and time information. The overall method uses a conditional diffusion model.

Training and testing details:

• The model was trained on hourly precipitation data from the ERA5 reanalysis (European Centre for Medium- Range Weather Forecasts) for 2020- 2023. Researchers simulated GSMaP- style gaps by removing corresponding areas from the complete ERA5 maps.
• On 2024 test data, it outperformed other deep- learning baselines (standard U-Net, variational autoencoder, and generative adversarial network models) in prediction accuracy and image quality while producing spatially and temporally consistent rainfall patterns.
• Sensitivity analysis showed that precipitation data and infrared images were the most important inputs.
• When applied to real incomplete GSMaP maps (using ERA5 as a reference, since no complete GSMaP ground truth exists), it successfully reconstructed the missing regions.

Paper details: “Conditional Diffusion Models for Global Precipitation Map Inpainting,” published in Earth and Space Science (Volume 13, Issue 9; available online 24 September 2026; DOI: 10.1029/2025EA004666).

The method could support near- real- time applications for heavy- rainfall and flood monitoring, water- resource management, hydrological analysis, and disaster- risk reduction, particularly valuable in regions with sparse ground gauges or radars. Funding included support from JAXA’s Precipitation Measuring Mission, JSPS KAKENHI grants, and Chiba University programs.

This work is distinct from earlier Chiba University research (e.g., by Muto and Kotsuki) that used the Local Ensemble Transform Kalman Filter (LETKF) to improve gauge- based global precipitation estimates.

Conditional Diffusion Models for Global Precipitation Map Inpainting

Abstract

Incomplete satellite- based precipitation data is a major challenge for global monitoring. Products such as JAXA’s Global Satellite Mapping of Precipitation (GSMaP) have large missing regions because polar- orbiting satellites carrying microwave sensors cannot observe the whole Earth at once. Existing interpolation methods often create spatial discontinuities.

The authors treat gap- filling as a video inpainting problem and introduce a conditional diffusion model. It uses a 3D U- Net plus a 3D condition encoder that incorporates infrared satellite images, latitude- longitude grids, and physical time information to reconstruct complete precipitation maps while preserving spatiotemporal consistency.

• Training data: ERA5 hourly precipitation (2020- 2023). Pseudo- GSMaP gaps were created by randomly applying real GSMaP observation masks to the complete ERA5 fields.
• Evaluation: Calendar year 2024.
• Result: The method produces more spatiotemporally consistent inpainted maps than conventional techniques and other deep- learning baselines.

The work demonstrates the potential of conditional diffusion models to improve continuous global precipitation monitoring.

Key Points

The approach outperforms conventional and other deep- learning methods at filling unobserved regions.

High-fidelity global precipitation maps are needed, but polar- orbiting satellites leave spatiotemporal gaps.

A diffusion- model inpainting method based on a 3D U- Net achieves strong spatiotemporal consistency.

Plain- Language Summary

Polar-orbiting satellites cannot see the entire planet at once, so large parts of global rainfall maps are missing. Current gap- filling methods mainly use neighboring hours and often produce unrealistic “seams” between observed and filled areas.

The team treats sequences of precipitation maps like video frames. A generative diffusion model learns to restore the missing parts by looking at nearby times, surrounding regions, and extra cues (infrared cloud images, geographic coordinates, and time).

Trained on complete reanalysis data with simulated satellite gaps (2020- 2023) and tested on 2024 data, the model fills gaps more smoothly and consistently than previous techniques.

Background and Motivation

Accurate, high- temporal- resolution global precipitation maps support extreme- weather mitigation, water- resource management, and policy.

Both GSMaP (JAXA) and IMERG (NASA) rely on microwave observations from polar- orbiting satellites and therefore contain large gaps. Conventional temporal interpolation prioritizes time consistency but neglects spatial consistency, producing discontinuities.

Method Overview

The approach is built on Denoising Diffusion Probabilistic Models (DDPM). A forward process gradually adds Gaussian noise; a reverse process learns to denoise and reconstruct the missing precipitation.

• Architecture: 3D U- Net (captures space + time) guided by a 3D condition encoder.
• Conditioning inputs: infrared cloud imagery, topographic data, latitude- longitude grids, and physical time.
• These multi- modal features are injected at multiple scales into the U- Net encoder blocks.

Training uses complete ERA5 fields with realistic GSMaP- style masks applied. The model is evaluated for accuracy, image quality, and spatiotemporal consistency against conventional interpolation and other deep-learning models (standard U-Net, VAE, GAN).

Significance

The method yields higher- fidelity, more consistent global precipitation maps and can support near- real- time applications such as heavy- rain and flood monitoring, hydrological analysis, and disaster- risk reduction, especially valuable in regions with sparse ground observations.

A free preprint with the full text, equations, architecture details, sensitivity analyses, and experimental results is available on arXiv. The published AGU version may contain minor updates or additional supporting information.

Journal: Earth and Space Science

DOI: 10.1029/2025ea004666

Provided: Chiba University, (AGU)

Authors: Daiko Kishikawa, Yuka Muto, Shunji Kotsuki

First published: 24 September 2026