Mask R-CNN based segmentation method for satellite imagery of photovoltaics generation systems

There is increasing interest in getting the precise locations and the corresponding sizes of installed photovoltaics. Previous and existing methods have been confirmed to be time-consuming, not robust or incomplete. Thus, we proposed a new method based on the satellite imagery, deep learning and image processing which can be used to collect the precise information of installed photovoltaic automatically. The information can be utilized to support the adoption and management of solar electricity. The method includes three main parts: the overlap-tile strategy, Mask R-CNN model and right-angle polygon fit algorithm. The overlap-tile strategy used here to improve the image edge segmentation ability of Mask R-CNN. The right-angle polygon fit algorithm is proposed to better fit the mask area generated by Mask R-CNN, which helps us get the much more precise locations and sizes of photovoltaics. The training and testing of our method were based on a satellite dataset with 3904 images annotated with ground truth regions of photovoltaics. The numerical and visual results clearly demonstrate that the accuracy and efficiency of our method is better than the previous reports.