High-resolution solar panel detection in Sfax, Tunisia: A UNet-Based approach

This study introduces a comprehensive method to enhance the precision of photovoltaic (PV) panel segmentation, crucial for the expanding utilization of PV technology in renewable energy. Conventional detection methods often lack precision, especially across varied resolutions. Leveraging recent advancements in machine learning techniques and image processing, we propose a UNet model-based approach to detect solar panels across different scales. Overcoming challenges posed by diverse sensing platforms and testing regions, the methodology demonstrates superior performance with reduced labeled datasets and computational requirements. Utilizing high-resolution aerial images and satellite data from Google Earth Pro, the model's training encompasses various zoom levels to bolster robustness. We enhance detection accuracy by incorporating three data enhancement methods: zoom-in, zoom-out, and blur. Tested in Sfax (eastern Tunisia), despite a relatively small dataset, our model achieves an impressive Intersection over Union (IoU) score of 86 %, enabling precise energy production estimation within the test area. Our framework identifies 499 photovoltaic (PV) systems covering 815.865 m2 in Sfax City, with an average regional PV implantation density of approximately 407 PV/km2 surpassing the national average. Notably, these PV installations are predominantly found on well-maintained villas, suggesting higher household income. To improve PV adoption rates, targeted promotion programs for lower-income households' electricity consumption can be effective in Tunisia. Our model can be used for accurate calculation of energy potential, thereby informing renewable energy planning and resource allocation strategies.