Yesterday, researchers from Stanford University introduced DeepSolar, a deep learning framework that analyzes satellite images to identify the GPS location and size of solar panels. Using this framework they have built a comprehensive database containing all the GPS locations and sizes of solar installations in the US. The system was able to identify 1.47 million individual solar installations across the United States, ranging from small rooftop configurations, solar farms, to utility-scale systems.
The DeepSolar database is available publicly to aid researchers to extract further insights into solar adoption. This database will also help policymakers in better understanding the correlation between solar deployment and socioeconomic factors such as household income, population density, and education level.
How DeepSolar works?
DeepSolar uses transfer learning to train a CNN classifier on 366,467 images. These images are sampled from over 50 cities/towns across the US with merely image-level labels indicating the presence or absence of panels.
One of the researchers, Rajagopal explained the model to Gizmodo, “The algorithm breaks satellite images into tiles. Each tile is processed by a deep neural net to produce a classification for each pixel in a tile. These classifications are combined together to detect if a system—or part of—is present in the tile.”
The deep neural net then identifies which tile is a solar panel. Once the training is complete, the network produces an activation map, which is also known as a heat map. The heat map outlines the panels, which can be used to obtain the size of each solar panel system.
Rajagopal further explained how this model gives better efficiency, “A rooftop PV system typically corresponds to multiple pixels. Thus even if each pixel classification is not perfect, when combined you get a dramatically improved classification. We give higher weights to false negatives to prevent them.”
What are some of the observations the researchers made?
To measure its classification performance the researchers defined two metrics: utilize precision and recall. Utilize precision is the rate of correct decisions among all positive decisions and recall is the ratio of correct decisions among all positive samples. DeepSolar was able to achieve a precision of 93.1% with a recall of 88.5% in residential areas and a precision of 93.7% with a recall of 90.5% in non-residential areas.
To measure its size estimation performance they calculated the mean relative error (MRE). It was recorded to be 3.0% for residential areas and 2.1% for non-residential areas for DeepSolar.
Future work
Currently, the DeepSolar database only covers the contiguous US region. The researchers are planning to expand its coverage to include all of North America, including remote areas with utility-scale solar, and non-contiguous US states. Ultimately, it will also cover other countries and regions of the world.
Also, DeepSolar only estimates the horizontal projection areas of solar panels from satellite imagery. In the future, it would be able to infer high-resolution roof orientation and tilt information from street view images. This will give a more accurate estimation of solar system size and solar power generation capacity.
To know more in detail, check out the research paper published by Ram Rajagopal et al: DeepSolar: A Machine Learning Framework to Efficiently Construct a Solar Deployment Database in the United States.