Deep Learning Framework for Multi-Class Segmentation of Photovoltaic Systems

Accurate mapping of photovoltaic (PV) installations supports renewable energy monitoring, capacity assessment, and policy planning. We introduce a deep learning framework for the simultaneous segmentation of rooftop (RTPV) and ground-mounted (GMPV) PV systems from 0.2m aerial imagery. The approach integrates automatically generated labels from registries and crowdsourced geographic information data with a small set of manually refined annotations, enabling scalable training with high spatial fidelity. A dual-head U-Net++ with a ResNet50 encoder jointly performs pixel-wise segmentation and image-level classification, reducing false positives in heterogeneous landscapes. Exemplarily applied to the state of Hesse, Germany, the model achieves a balanced accuracy of 87.52%, precision of 94.74%, and a recall of 82.89%, with F1-scores of 75.75% (RTPV) and 88.65% (GMPV). Coupled with solar suitability data, the results indicate rooftop PV utilization of only 3.24% statewide and 2.5% in the Dietzenbach case study. This implemented deep learning framework is the first high-resolution, single-step multi-class PV segmentation applied at the federal state scale, offering a transferable tool for large-area PV monitoring.