The expansion of photovoltaic (PV) installations is crucial for global energy transition, but detailed information on their spatial distribution remains scarce, posing challenges for effective energy planning. This study presents a methodology for the automatic recognition of ground-mounted PV systems in Italy, using semantic segmentation and Sentinel-2 10-meter-resolution RGB images. The proposed methodology aims at accurately detecting both locations and sizes of plants, estimating capacity and ensuring regular map updates, to support energy planning strategies.
The segmentation model, based on a U-Net architecture, is trained on a dataset from 2019 and tested on two distinct cases, involving different imagery dates and areas. We propose a multi-temporal approach, applying the model to a series of images captured throughout the year and aggregating outputs to create a PV detection probability map. Users can adjust probability thresholds to optimize accuracy: lower thresholds enhance Producer Accuracy, ensuring continuous area detection to estimate capacity, while higher thresholds improve User Accuracy by minimizing false positives. Post-processing methods, such as plastic-covered greenhouse filtering, help reduce detection errors. Nonetheless, model generalizability across diverse landscapes needs improvement, requiring retraining with images from various environmental contexts.