Photovoltaic Solar Array Mapping using Supervised Fully Convolutional Neural Networks

Abstract

This study explores a supervised deep learning fully convolutional segmentation model for photovoltaic solar array mapping from aerial imagery. The deep learning imaging techniques present a fast and an inexpensive way for detecting distributed photovoltaic arrays installed on ground and building rooftops. The identification of correct photovoltaic array shapes and sizes is a necessary requirement for the estimation of energy from photovoltaic arrays within an area or city. This study proposes a modified and efficient UNet deep learning segmentation model by using depthwise-separable convolution for automated photovoltaic array detection from orthorectified RGB imagery with a resolution of less or equal to 0.3m. The result shows our model has better segmentation accuracy than various state of the art models and other previous studies on solar panel detection and is efficient in terms of parameters and complexity.