Researchers at Stanford University have managed to create a map of almost every solar rooftop in America, pulling together satellite images and extracting the information. The DeepSolar data, based on Google’s Inception V3, is the first publicly available, high-fidelity solar installation database in the contiguous United States, the researchers say.
A total of 1.47 million solar rooftops were identified in the lower 48 states in the study. “We built a nearly complete solar installation database for the contiguous United States utilizing a novel deep learning model applied to satellite imagery,” the researchers say. They published the method and results of their study in Joule, Volume 2, Issue 12, 19 December 2018.
DeepSolar learned to identify solar panels by analyzing some 370,000 images representing areas of 100 feet by 100 feet. Then DeepSolar learned to identify features associated with solar panels including roof color, texture, and size. The system was able to correctly identify 93% of solar rooftops, according to a Stanford report on the project.
The DeepSolar team included Jiafan Yu, Zhecheng Wang, Arun Majumdar and Ram Rajagopal. Work was also contributed by the Majumdar Group’s Stanford-based Magic Lab, and the Stanford Sustainable Systems Lab, in Palo Alto, CA. Funding for the project included a State Grid Fellowship from Stanford Energy’s Bits & Watts initiative, and a Stanford Interdisciplinary Graduate Fellowship.
“We developed DeepSolar, a deep learning framework analyzing satellite imagery to identify the GPS locations and sizes of solar photovoltaic panels. Leveraging its high accuracy and scalability, we constructed a comprehensive high-fidelity solar deployment database for the contiguous United States, the authors say.
Google’s Inception V3 is “a widely-used image recognition model that has been shown to attain greater than 78.1% accuracy on the ImageNet dataset. The model is the culmination of many ideas developed by multiple researchers over the years,” Google states.
“The recent breakthroughs of deep learning enables automatic and accurate image classification and segmentation. Combining satellite imagery and deep learning, we aimed to develop a framework to automatically construct, maintain, and update the solar installation database and realize the next-level visibility on renewable energy deployment,” the team said.
The team plans to add features to calculate a solar installation angle and orientation, which could accurately estimate its power generation. Plans also are to update the map annually and add other countries and regions of the world.
“We offer the DeepSolar database as a publicly available resource for researchers, utilities, solar developers, and policymakers to further uncover solar deployment patterns, build comprehensive economic and behavioral models, and ultimately support the adoption and management of solar electricity,” the team said.
Apart from the novelty of the feat, the researchers identified key socioeconomic factors correlating with solar deployment density, they say. They discovered that residential solar deployment density peaks at a population density of 1,000 per capita per square mile, and that it increases rapidly with annual household income starting at $150,000. “Furthermore, we built an accurate machine learning-based predictive model to estimate the solar deployment density at the census-tract level,” the team stated.
The study also revealed that with a solar radiation reception threshold of 4.5 kilowatt-hours per square meter per day, solar deployment is “triggered.”
Data used in the study came from Google, and was “pre-trained” on Stanford’s ImageNet image database. “DeepSolar model incorporates both image classification and semantic segmentation in a single Convolutional Neural Network. Classification is to localize the solar panels and segmentation is to estimate their sizes. The classification branch is developed based on Google Inception V3, which is pretrained on ImageNet and then fine-tuned on our dataset containing 360K images,” the team explains. “The output of the classification branch is a class indicating either “positive” (containing solar panel) or “negative” (not containing solar panel). The precision and recall of classification are both around 90% for residential and non-residential areas,” they add.
Earlier solar maps of the United States often are depictions of the total amount of sunshine that arrives on a given spot per day. A map commonly used in the industry is the State University of New York at Albany’s satellite radiation model, developed by Richard Perez and collaborators at the National Renewable Energy Laboratory and other universities for the US Department of Energy.
This SUNY model uses hourly radiance images from geostationary weather satellites, daily snow cover data, and monthly averages of atmospheric water vapor, trace gases, and the amount of aerosols in the atmosphere to calculate the hourly total insolation from both the sun and the wider sky falling on a horizontal surface. Atmospheric water vapor, trace gases, and aerosols are derived from a variety of sources.