Global rainfall maps are built from observations collected by multiple satellites, but they have an important limitation. The satellites that provide these measurements cannot see the whole Earth at once. As a result, some regions of these maps can have no valid observations.
Current methods try to fill these gaps by estimating precipitation values from observations made before and after the missing period. However, these methods may not fully capture how precipitation patterns change over time, resulting in inconsistencies in the maps.
To address this challenge, researchers led by Assistant Professor Daiko Kishikawa from the Institute for Advanced Academic Research / Center for Environmental Remote Sensing , Chiba University, Japan, treated the problem as a type of video reconstruction task. In this approach, the team estimates missing parts of a sequence using information from surrounding frames.
The research team, which included Dr. Yuka Muto from the Center for Environmental Remote Sensing, Chiba University, and Dr. Shunji Kotsuki from the Institute for Advanced Academic Research/Center for Environmental Remote Sensing/Research Institute of Disaster Medicine, Chiba University, proposed a machine learning approach that learns how precipitation patterns change across both space and time and estimates how rainfall would be distributed in the missing areas.
This study was made available online on September 24, 2026, and published in Volume 13, Issue 9 of the journal Earth and Space Science on September 01, 2026.
"Instead of relying on manually designed formulas, our method treats precipitation map sequences as video-like data and repairs missing sections by using information from both nearby time frames and the surrounding regions with additional information," explains Dr. Kishikawa.
The approach uses a three-dimensional (3D) U-Net, a convolutional neural network architecture commonly used for image segmentation, to capture precipitation patterns across space and time. The U-Net is combined with a 3D condition encoder that processes additional information, including infrared satellite images, geographic coordinates, and time, to help reconstruct complete precipitation maps.
Their aim was to reconstruct missing observations in the Global Satellite Mapping of Precipitation (GSMaP), produced by the Japan Aerospace Exploration Agency.
To train the model, the researchers used hourly precipitation data from the ERA5 reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts, covering the period from 2020 to 2023. ERA5 provides complete precipitation maps, so the researchers used the locations of missing observations in GSMaP to remove the corresponding areas from the ERA5 maps. These pseudo-GSMaP maps resembled incomplete satellite observations and allowed the researchers to train the model to reconstruct the missing regions.
The proposed approach uses a conditional diffusion model to reconstruct the missing precipitation information. The 3D condition encoder processes this information and provides the U-Net with the details it needs to reconstruct the missing precipitation.
When tested on 2024 data, the model achieved high prediction accuracy and image quality compared with other deep learning approaches, including U-Net, variational autoencoder, and generative adversarial network models. It also produced rainfall patterns that remained consistent across locations and over time. Sensitivity analysis showed that precipitation data and infrared satellite images were the most important inputs for accurately reconstructing missing precipitation.
To see how the model would perform in practice, the researchers applied it to real GSMaP maps with missing regions. Since a complete GSMaP dataset did not exist for comparison, the researchers used ERA5 data as a reference instead. The results showed that the model could reconstruct missing regions in real GSMaP data, demonstrating its potential for application to actual satellite observations.
The approach could also be developed for near-real-time applications, providing more complete rainfall information for monitoring heavy rainfall, managing water resources, and reducing disaster risks.
"This could contribute to a wide range of applications that rely on precipitation information, including heavy-rainfall and flood monitoring, water-resource management, hydrological analysis, and disaster risk reduction," says Dr. Kishikawa.
By making global rainfall information more continuous, such an approach could be valuable in regions where observations are sparse, such as those with few ground-based rain gauges or weather radars.
To see more news from Chiba University, click here .
Reference:
Title of original paper: Conditional Diffusion Models for Global Precipitation Map Inpainting
Authors: Daiko Kishikawa1,2, Yuka Muto1, and Shunji Kotsuki1,2,3,4
Affiliations:
1Center for Environmental Remote Sensing, Chiba University, Japan
2Institute for Advanced Academic Research, Chiba University, Japan
3Research Institute of Disaster Medicine, Chiba University, Japan
4Data Assimilation Research Team, RIKEN Center for Computational Science, Japan
Journal: Earth and Space Science
DOI: https://doi.org/10.1029/2025EA004666
About Assistant Professor Daiko Kishikawa from Chiba University, Japan
Dr. Daiko Kishikawa is a Specially Appointed Assistant Professor at the Institute for Advanced Academic Research at Chiba University, Japan. He earned his Ph.D. in Engineering from Chiba University in 2023 and is currently affiliated with the Kotsuki & Okazaki Lab at the Center for Environmental Remote Sensing (CEReS). His research focuses on artificial intelligence, specifically Inverse Reinforcement Learning, and Deep Reinforcement Learning. He applies these methodologies to diverse fields, ranging from autonomous driving comfort and weather prediction modeling to data-driven analysis of classical Japanese literature in digital humanities.
Funding:
This study was partially supported by the Japan Aerospace Exploration Agency (JAXA) Precipitation Measuring Mission, JSPS KAKENHI (Grants JP21H05002, JP2SKI7687 and JP2SHO0752), and the IAAR Research Support Program and VL Program of Chiba University.
