Solar panel detection in Dutch municipalities using satellite images

Abstract

Photovoltaic (PV) and Solar Thermal (ST) panels mounted on rooftops form a cornerstone in the transition to fully renewable energy generation. However, due to the large gap in data on the number and location of these panels, policymakers have trouble determining the effectiveness of policies and energy network administrators have trouble building efficient networks. In this study, a model is proposed to automatically classify and segment PV and ST panels from aerial imagery to alleviate this issue. A novel dataset of aerial images in the Netherlands, containing image-level and pixel-level annotations of PV and ST panels is presented and made publicly available. A two-stage pipeline consisting of a classification and segmentation stage is proposed, as well as a novel method for weakly-supervised pseudo-label generation based on greedy Class Activation Map (CAM) refinement and Segment Anything Model (SAM) generated segmentations. The model is shown to exhibit strong classification performance, after finetuning models pretrained either on ImageNet or Dutch aerial images. Performance of fully-, semi-, and weakly-supervised segmentation models is evaluated. It is shown that the best performance is achieved by combining a small set of manually annotated mask labels with a larger set of unlabelled data in a semi-supervised manner. This semi-supervised approach leads to an IoU of 73.3% for binary segmentation, and a class-specific IoU of 77.0% and 37.6% is achieved for the PV and ST classes respectively.