SolarDetector: Automatic Solar PV Array Identification using Big Satellite Imagery Data

Due to the intermittent nature of solar energy, it has been increasingly challenging for the utilities, third-parties, and government agencies to integrate distributed energy resources generated by rooftop solar photovoltaic (PV) arrays into smart grids. Recently, there is a rising interest in automatically collecting solar installation information in a geospatial region that are necessary to manage this stochastic green energy, including the quantity and locations of solar PV deployments, and their profiling information. Most recent work focuses on using big aerial or satellite imagery data to train machine learning or deep learning models to automatically detect solar PV arrays. Unfortunately, these approaches are suffering low detection accuracy due to the insufficient sample and feature learning when building their models, and the separation of rooftop object segmentation and identification during their detection process. In addition, most recent approaches cannot report accurate multi-panel detection results. To address these problems, we design a new approach---SolarDetector that can automatically detect and profile distributed solar photovoltaic arrays in a given geospatial region without any extra cost. SolarDetector first leverages data augmentation techniques and Generative adversarial networks (GANs) to automatically learn accurate features for rooftop objects. Then, SolarDetector employs Mask R-CNN algorithm to accurately identify rooftop solar arrays and also learn the detailed installation information for each solar array simultaneously.

This presentation is part of the 2023 Energy Data Analytics Symposium: Accelerating Sustainability in the AI Era: https://nicholasinstitute.duke.edu/ev...

Learn about the Energy Data Analytics Lab at Duke's Nicholas Institute for Energy, Environment & Sustainability: http://nicholasinstitute.duke.edu/iss...

Get email updates from the Nicholas Institute: http://nicholasinstitute.duke.edu/new...