Analysis of Building Segmentation Method for Calculating Photovoltaic Potential Area Using Segment Anything Model (SAM) on the Massachusetts Building Dataset

Accurate roof segmentation in aerial images is essential for precise calculation of photovoltaic potential, which is increasingly important for optimizing solar energy utilization. This study aims to enhance roof segmentation precision by developing an architecture based on the Segment Anything Model (SAM) and comparing its performance to a baseline UNET architecture. We evaluated the inherent differences between instance segmentation (SAM) and semantic segmentation (UNET) approaches. The predicted roof areas of the proposed SAM-based instance segmentation model were quite accurate, hence masking those areas was facilitated. Some of the performance metrics include accuracy, Dice coefficient, and IoU. With the SAM model, an accuracy of 85.48%, a Dice coefficient of 57%, an IoU of 40.4%, while UNET established an accuracy of 84.89%, a Dice coefficient of 84.7<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, and an IoU of 73.6% accordingly. In fact, these results really show that while the SAM model is more accurate, the UNET model outperforms it in terms of Dice coefficient and IoU, probably because of fundamental differences between instance and semantic segmentation approaches. As conclusion, instance segmentation models such as SAM have shown their great potential in detecting roofs, whereas semantic segmentation models like UNET perform better in detecting them. A web application was developed using Streamlit and FastAPI to visualize segmentation results, effectively mapping building roofs but with room for improvement.