Automatic Estimation of Solar Rooftops and Power Generation From Publicly Available Satellite Imagery Through Georeferencing and Large-Scale Support

Abstract

Rooftop photovoltaic (PV) power systems constitute a viable alternative energy technology that can significantly reduce electricity costs. The rapid increase in installations has led to a mismatch between planned power generation and actual electricity demand, necessitating effective monitoring and impact assessment. This study proposes a novel approach for detecting solar rooftops using publicly available satellite imagery over large areas. We also introduce a technique for estimating solar panel size and potential energy production, with outputs formatted for GIS applications. Employing a modified U-Net architecture with pre- and post-processing techniques, our experiments achieved an Intersection over Union score of 0.7879 and a Dice score of 0.8808. Image tiling and mosaicking with georeferencing were used to support large-scale imagery. The detection results were post-processed through polygonization and smoothing using the Douglas-Peucker algorithm. Panel size and power generation were then calculated and attached as attributes. Through satellite image analysis, this study aims to accurately identify and evaluate solar rooftops nationwide, providing valuable insights for homeowners, businesses, and government authorities. This facilitates informed decision-making, cost reduction, and contributions to environmental goals.

Keywords

• Attribute extraction
• deep learning
• satellite images
• semantic segmentation
• solar panel