Leveraging Semantic Segmentation for Photovoltaic Plants Mapping in Optimized Energy Planning

The growth of photovoltaic (PV) installations is essential for the global energy transition; however, comprehensive data regarding their spatial distribution are limited, which complicates effective energy planning. This research introduces a methodology for automatic recognition of ground-mounted PV systems in Italy, using semantic segmentation and Sentinel-2 RGB images with a resolution of 10 m. The objective of this methodology is to accurately identify both the locations and the sizes of these installations, estimate their capacity, and facilitate regular updates to maps, thereby supporting energy planning strategies. The segmentation model, which is founded on a U-Net architecture, is trained using a dataset from 2019 and evaluated on two separate cases that involve different dates and geographical areas. We propose a multi-temporal approach, applying the model to a sequence of images taken throughout the year and aggregating the results to create a PV detection probability map. Users have the flexibility to modify probability thresholds to enhance accuracy: lower thresholds increase producer accuracy, ensuring continuous area detection for capacity estimation, while higher thresholds boost user accuracy by reducing false positives. Additionally, post-processing techniques, such as filtering for plastic-covered greenhouses, assist minimizing detection errors. However, there is a need for improved model generalizability across various landscapes, necessitating retraining with images from a range of environmental contexts.