CONVOLUTIONAL NEURAL NETWORK FOR DETECTION OF RESIDENTIAL PHOTOVOLTAIC SYSTEMS IN SATELLITE IMAGERY

Due to recently growing adoption, residential photovoltaic (PV) installations are becoming a key contribution to renewable energy production. In order to efficiently track the inherently decentralized deployment of these PV systems, this study leverages widely available high-resolution satellite imagery, along with the demonstrated ability of convolutional neural networks (CNNs) in image classification tasks. In particular, this effort presents development of a custom-trained CNN that operates on images of residential areas received as part of a larger image processing pipeline. This custom-trained CNN is shown to achieve comparable accuracy to the state-of-the-art achieved via transfer learning, but with reduced computational burden and required image resolution. Using imagery from two cities in California, this approach achieves PV classification with a precision of 91.9% and recall of 92.4%. This accuracy is sufficient to inform actionable insights from satellite imagery regarding the political, social, and economic factors affecting PV deployment.