Key points are not available for this paper at this time.
Timely extraction of high-quality photovoltaic (PV) panels from high-resolution remote sensing imagery can contribute to a comprehensive understanding of energy production, and supports global carbon neutrality and Sustainable Development Goals (SDGs). Existing studies for extracting PV panels only focus on small-scale rooftop PV systems but not on large-scale PV systems composed of single or several PV panels or arrays. Furthermore, the currently available public datasets related to large-scale PV systems are limited by the coarse resolution of the imagery used or annotation scale and do not provide highly detailed footprints for PV panels or attributes such as the location, quantity, or area of these panels. To fill this gap, a novel semantic segmentation model (PVNet) for extracting high-quality PV panels from the densely distributed and regularly shaped PV panels in large-scale PV systems is proposed. PVNet consists of two modules, a Coarse Prediction Module (CPM) and a Fine Optimization Module (FOM). The CPM extracts complete regions of individual PV panels by fusing low-level local features with high-level global features, while the FOM optimizes the output of CPM by residual refinement to match extracted boundaries to ground truth. High-quality details from region to boundary of PV panels are obtained through joint supervision of CPM and FOM results. PVNet was trained on a newly annotated PV Panel Dataset and tested under four scenario conditions in China where the large-scale PV industry is growing rapidly. Qualitative and quantitative results show that PVNet can achieve the highest accuracy for PV panel extraction with F1socre higher than 0.88 and IoU higher than 0.79. Area-wide PV panel mapping and comparisons with existing PV footprint datasets demonstrate that PVNet is a feasible solution for obtaining high-quality geo-spatial databases of large-scale PV systems.
Wang et al. (Sat,) studied this question.