AI-enabled rooftop solar mapping and optimization: A computer vision approach for photovoltaic systems

The rapid growth in energy and environmental issues, rooftop PV systems are rapidly gaining popularity. But one of the problems that needs to be addressed in large-scale deployment of solar is the problem of identifying appropriate rooftops accurately and in a scalable way. In this study, an integrated AI vision system is proposed that uses YOLOv8 and the Segment Anything Model (SAM) to detect and segment rooftops in high-resolution satellite imagery. YOLOv8 was optimized on a dataset with 70% training, 15% validation, and 15% test data of urban rooftops from the cities of Pakistan, mainly Karachi and Lahore, and was trained for 300 epochs with the AdamW optimizer. SAM was then used to refine the YOLOv8-detected bounding boxes to produce accurate pixel-wise segmentation masks. The proposed approach outperforms current methods such as U-Net, Mask R-CNN, and Deep Roof with a detection precision of 92.8%, recall of 90.7%, mAP@0.5 of 94.1%, and SAM segmentation IoU of 89.3%. To estimate PV installation capacity, segmented roof areas were estimated using geospatial metadata. The findings show that the YOLOv8-SAM pipeline is a repeatable and scalable solution to manual surveying of rooftops and can be used to aid in solar energy planning in urban areas where data is limited.