The solar industry is growing steadily in the US, with revenues ballooning from US$42 million in 2007 to US$210 million in 2017, while total capacity is expected to more than double over the next five years. Though these broad strokes paint a useful overall picture of rates of solar uptake, there's a lot to be learned from drilling into the finer details. Stanford scientists have built a new machine learning tool specifically to do the job, and it goes by the name of DeepSolar.

Knowing where solar panels are and what motivated folks to install them may prove invaluable to energy management efforts. It could help utility companies better balance supply and demand, and therefore provide more reliable power. It can also help us understand what inspires people to install rooftop solar and others not to, and perhaps design cities accordingly.

As it stands, researchers can only go by rough estimates, but with satellite imagery improving all the time new possibilities have emerged. The Stanford scientists trained a machine learning algorithm to tackle this momentous task by feeding it around 370,000 images, each presenting an area of Earth covering around 100 ft x 100 ft (30 x 30 m) and each labeled as to whether or not they contained a solar panel.

By analyzing these images, the DeepSolar program determined the types of features, such as colors, textures and sizes, that could reliably be associated with a solar panel. With time, DeepSolar got quite good at this, and was able to correctly identify images with solar panels 93 percent of the time, though it missed around one in 10 that did.

"We don't actually tell the machine which visual feature is important," says Jiafan Yu, a Stanford doctoral candidate in electrical engineering who built the system with Zhecheng Wang, a doctoral candidate in civil and environmental engineering. "All of these need to be learned by the machine."

The team then put DeepSolar to work analyzing a billion satellite images in search of US solar installations, which took it just one month. It unearthed solar panels on residential properties, the roofs of business and at large solar plants, totaling 1.47 million installations in all, a figure that far exceeds current estimates. This was integrated with US Census and other data to draw conclusions about motivating factors behind solar power adoption.

For example, the team found that low and medium income homes often won't install solar, even in areas with lots of sunshine where it would be profitable in time, which they suspect is due to the upfront costs. Another interesting tidbit came through the integration of geographic data that enabled the team to identify a threshold of sunlight exposure needed to trigger adoption in a given area.

"We found some insights, but it's just the tip of the iceberg of what we think other researchers, utilities, solar developers and policymakers can further uncover," says supervisor Arun Majumdar. "We are making this public so that others find solar deployment patterns, and build economic and behavioral models."

All of the DeepSolar data is publicly available on the project's website, while the research has been published in the journal Joule.