Forget Cats, This Neural Network Spots Solar Panels

Stanford’s DeepSolar neural network analyzed satellite images to count U.S. solar installations—and there are a lot more than anybody thought

2 min read

Satellite image with solar panels highlighted with yellow circles.
Image: DeepSolar/Stanford

There are at least 1.47 million solar installations of varying sizes in the 48 contiguous U.S. states, from home rooftop panels to utility-owned solar power plants.

That’s the conclusion of DeepSolar, a machine learning algorithm developed by researchers at Stanford University that searches satellite images for solar panels. The count is higher than some previous estimates, like the OpenPV project’s count of 1.02 million installations.

The researchers, led by Ram Rajagopal, associate professor of civil and environmental engineering, and Arun Majumdar, professor of mechanical engineering, trained DeepSolar on a set of 370,000 satellite images, each covering a region measuring approximately 9 square meters (100 square feet), by indicating which ones included solar panels. The machine-learning program then figured out how to identify solar panels, spotting them correctly 93 percent of the time. When it erred, it tended to undercount the installations, missing about 10 percent.

Selection of satellite images to show the deep learning algorithm identifying solar panels.The DeepSolar deep-learning framework analyzes satellite imagery to identify the GPS locations and sizes of solar photovoltaic panels.Image: DeepSolar/Stanford

Processing the full set of a billion images took about a month. For efficiency, the system did not review the most sparsely populated areas of the United States, which the researchers estimate would add about 5 percent to the overall count. They plan to add those regions to future runs of the program, and also aim to calculate the angle and orientation of the panels to more accurately estimate power generation.

As part of this analysis, the researchers created a database of the solar installations and added in U.S. census data. They found that low- and medium-income households do not often install solar systems even when located in areas with high electric bills and lots of sunshine. They also discovered that there is a solar radiation threshold—4.5 kWh/m2/day—that triggers solar adoption, although income levels can affect that. Using this data, DeepSolar can now predict solar deployment density for a particular area.

The researchers released their results today to coincide with the publication of a paper in Joule. More information along with the raw data and an interactive map is available here. They plan to update the count annually.

The Conversation (0)