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

Abstract

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. In addition, SolarDetector could also integrate with large-scale data processing engine - Apache Spark and graphics processing units (GPUs) to further improve its training cost. We evaluate SolarDetector using 263,430 public satellite images from 11 geospatial regions in the U.S. We find that pre-trained SolarDetector yields an average MCC of 0.76 to detect solar PV arrays over two big datasets, which is ∼50% better than the most notable approach - SolarFinder. In addition, unlike prior work, we show that SolarDetector can also accurately report the profiling information for the detected rooftop objects.