AI Sharpens Rainfall Estimates from Satellites

Better precipitation data can improve climate change models

3 min read
a cloud made of lines connecting with sparkle lines below it
iStock

This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore.

Weather satellites can track every movement of a storm from orbit. However, the algorithms that analyze satellite readings currently don’t estimate precipitation levels as well as the algorithms that have been fed ground-based radar data. Now, a new study finds AI can help greatly improve satellite-based estimates.

Accurate estimates of how much rain, snow and hail has fallen on an area is important to refining weather forecasts, managing vital water supplies, and investigating climate change. But developing accurate precipitation estimates from raw satellite data “is not an easy task,” explains study co-author Haonan Chen, a remote sensing scientist at Colorado State University in Fort Collins. Satellites essentially scan cloud tops instead of detecting what has fallen on the surface. “Extracting surface-level precipitation information from measurements on top of the clouds requires sophisticated models that can bridge the gap,” he says.

In the new study, Chen and his colleagues analyzed data from the Geostationary Operational Environmental Satellites (GOES) operated by the U.S. National Oceanic and Atmospheric Administration. The researchers focused on the GOES-R satellites, which scan visible and infrared light from Earth.

“Combining the geostationary satellite data shared from other countries in Asia, we can potentially create a complete precipitation map covering the whole globe.” —Haonan Chen, Colorado State University

The researchers analyzed GOES-R data using deep learning techniques, which employ artificial neural networks. In these systems, components dubbed “neurons” are fed data and cooperate to solve a problem, such as recognizing images. The neural net repeatedly adjusts the links between its neurons and sees if the resulting patterns of behavior are better at finding a solution. Over time, the network discovers which patterns are best at computing results. It then adopts these as defaults, mimicking the process of learning in the human brain.

The aspects of a neural net that change with learning, such as the nature of the connections between neurons, are known as its parameters. A neural network is dubbed “deep” if it possesses multiple layers of neurons.

The scientists experimented with a neural network with more than 1.3 million parameters. They had it analyze GOES-R infrared data for the southwestern United States from May to September 2018 and May to September 2019, training it to get precipitation estimates as close as possible to ground-based radar estimates. They also investigated how their system performed when it did and did not incorporate data from GOES-16’s lightning detector.

“Thanks to GPU servers available in recent years, we can now analyze large amounts of satellite data and train complex machine learning models in a rather efficient way,” Chen says.

When tested on GOES-R data from May to September 2020, the researchers found their AI system was significantly better at matching ground-based radar estimates than other algorithms used to analyze satellite data. They also found when their AI model incorporated lightning data, it proved significantly more accurate in estimating heavy precipitation than when it did not.

Although this new study focused on the southeastern United States, the GOES-R satellites cover the western hemisphere, “so this work can be extended to a much larger domain,” Chen says. “In fact, combining the geostationary satellite data shared from other countries in Asia, we can potentially create a complete precipitation map covering the whole globe.”This means that researchers could benefit from AI-augmented precipitation estimates in places where they lack ground-based weather radar.

Future research will focus on testing how well these AI systems can generalize to analyzing data from other locations, Chen says. Giving the AI models even more data—such as readings concerning wind, moisture, and local terrain features—may improve their estimates even further, he adds.

A potential obstacle they face is how to handle the small-scale variability in precipitation often seen in complex terrains. “Unfortunately, satellites often face challenges observing such variability,” Chen says.

The scientists detailed their findings 8 September in the journal IEEE Geoscience and Remote Sensing Letters.

The Conversation (0)