Machine learning video-inpainting method fills gaps in global satellite rainfall maps

October 20, 2026

A research team led by Chiba University in Japan has developed a machine learning model that fills missing parts of global rainfall maps by treating them like missing sections of a video

Published in Earth and Space Science, the study shows that a type of AI model called a conditional diffusion model can accurately fill in missing areas in satellite rainfall data. This creates global rainfall maps that look more complete and connected.

Using video repair methods to fill gaps in satellite data

Global rainfall datasets, such as the Japan Aerospace Exploration Agency's (JAXA) Global Satellite Mapping of Precipitation (GSMaP), use microwave radiometers carried by satellites in low Earth orbit (LEO). These satellites cannot observe the whole Earth at the same time, so their raw measurements often contain large areas with no data.

Traditional mathematical methods estimate missing values using nearby frames from different times. However, they often struggle to keep rainfall patterns connected across an image, leaving visible marks along the edges of filled areas.

To address this problem, Assistant Professor Daiko Kishikawa and his colleagues treated the missing data as a video-filling task:

1. 3D U-Net Architecture:
   • The model uses a three-dimensional convolutional neural network called a 3D U-Net. It processes several frames together to learn how rainfall changes across places and over time.
2. Multimodal Condition Encoder:
   • A 3D condition encoder sends extra information directly into the U-Net to help it make predictions. This information includes infrared satellite images, grids showing latitude and longitude, and information about the time of each observation.
3. Conditional Diffusion Framework:
   • The model gradually removes noise from its predictions. This helps it create realistic rainfall patterns while accounting for uncertainty and natural variation.

How the model was trained and tested

The researchers trained the model using hourly global rainfall data from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis dataset, covering 2020 to 2023. To create examples of missing data, they placed masks based on real gaps in GSMaP observations over complete ERA5 rainfall maps.

The researchers tested the model using data from 2024. The conditional diffusion method performed better than several standard deep learning methods, including ordinary U-Nets, Variational Autoencoders (VAEs), and Generative Adversarial Networks (GANs):

1. Consistency across places and time:
   • Tests showed that using rainfall sequences together with infrared satellite images helped the model keep storm shapes connected at their edges and maintain smooth changes from one time to the next.
2. Use with real satellite data:
   • When the researchers used the model on actual GSMaP maps with missing data, it filled the gaps successfully. The resulting rainfall maps were continuous and matched the reference datasets.

How this could help with disaster planning

Being able to create complete rainfall maps almost in real time is especially important in places without ground-based weather radar networks or rain gauges.

The authors say that adding this conditional diffusion method to working satellite monitoring systems could greatly improve flood forecasts, early warning systems, models of how water moves through the environment, and water resource management.