Automatic solar panel detection from high-resolution orthoimagery using deep learning segmentation networks

Abstract

Solar panel detection from aerial or satellite imagery is a very convenient and economical technique for counting the number of solar panels on the rooftops in a region or city and also for estimating the solar potential of the installed solar panels. Detection of accurate shapes and sizes of solar panels is a prerequisite for successful capacity and energy generation estimation from solar panels over a region or a city. Such an approach is helpful for the government to build policies to integrate solar panels installed at home, offices, and buildings with the electric grids. This study explores the use of various deep learning segmentation algorithms for automatic solar panel detection from high-resolution ortho-rectified RGB imagery with resolution of 0.3 m. We compare and evaluate the performance of six deep learning segmentation networks in automatic detection of the distributed solar panel arrays from satellite imagery. The networks are tested on real data and augmented data. Results indicate that deep learning segmentation networks work well for automatic solar panel detection from high-resolution orthoimagery.