Solar or Not Solar: Lumen Energy Makes Strides in Image Classification

Environmentalists and industrialists alike can finally put aside their differences and appreciate a new model developed by UC Berkeley students for Lumen Energy, a San Francisco Bay Area-based startup. This model can be used to identify the presence of solar panels in satellite rooftop imagery with 98% accuracy, allowing Lumen to determine which buildings already have solar installations and thus estimate the current level of solar penetration for any given geographic area.

Steven Banjeree, a bioengineer and visiting scholar at UC Berkeley, and Peter Light, a former Googler with extensive product management experience in the energy industry, teamed up to create Lumen with the ultimate goal of unlocking the rapid deployment of solar technology in cities across California and beyond. “Here at Lumen Energy, we cater to a variety of stakeholders. This model is another step towards achieving our goals of making it easier for building owners to get paid to rent their roofs, solar installers to discover lucrative projects, and investors to fund low-risk, high-return clean energy projects,” said Banerjee.

To create this model, Banjeree and Light teamed up with students from Data-X, an applied data science course offered by the Sutardja Center for Entrepreneurship and Technology at UC Berkeley. Working together, the team came up with a Convolutional Neural Network binary classification model that was trained on over 16,000 satellite images. Their model achieved tremendous results, yielding 98% accuracy, 93% precision, and 94% recall on a test set of over 800 satellite images.

To demonstrate the power of this model, Lumen has created an interactive 3D map that allows users to visualize where solar installations are located within Hayward, California. This map includes classifications for over 50,000 buildings in the city, identifying the presence of solar panels in around 6.5% of those analyzed. Going forward, Lumen plans to expand the scope of their analysis to include buildings across the Bay Area and the entire state of California.

Project by: Jessica Houghton, Gordon Hu, Eric Phillips, Padma Teja