The automatic segmentation of residential solar panels based on satellite images: A cross learning driven U-Net method

Abstract Segmenting small-scale residential solar panels (RSPs) based on satellite images is an emerging data science problem in the renewable energy field. In this paper, we develop a cross learning driven U-Net (CrossNets) method and its extension, adaptive CrossNets, to automatically segment RSPs in satellite images. Proposed methods employ a group of generic U-Nets as a community and target to enhance the RSP segmentation performance. First, parameters of each generic U-Net in the community of CrossNets are initialized individually via the initialization with transfer learning and the classical initialization methods. Next, a novel training mechanism, cross learning, is developed to serve as a constraint for better optimizing CrossNets. Based on cross learning, each generic U-Net in the community first individually updates parameters at every epoch and next learns parameters from the best individual at specific epochs. Cross learning relieves the reliance of generic U-Nets on a careful initialization and better optimizes U-Nets. In testing, the result of the best performed generic U-Net in the community is selected as the final segmentation result of CrossNets. Adaptive CrossNets, a variant of CrossNets, is developed by applying an additional threshold to reduce the possibility of over-learning caused by cross learning. Satellite images collected from one city in U.S. are utilized to validate the performance of proposed methods. These images cover a large area of 135 km2 with 2794 RSPs. Compared with two generic U-Nets based benchmarks, our method can enhance the overall segmentation IoU by around 34% and 1.5%. Moreover, the segmentation robustness is improved from 1.191e−2 and 1.286e−4 to 2.481e−5. In addition, two new image datasets collected from other two cities in U.S. are applied to further examine the applicability of proposed methods.