Convolutional Neural Network-Based Roof Classification and Segmentation for Solar Potential Estimation

Abstract

This study presents a convolutional neural network-based pipeline for automated roof classification and segmentation to estimate residential solar potential using satellite imagery from Google Earth, BRAILS, and Kaggle. A sequential convolutional neural network (CNN) accurately classified roof types (flat, gabled, hipped) at 90.63%, and an improved U-Net segmented rooftop areas with 89.03% accuracy. Segmentation performance declined for white-colored roofs, emphasizing the need for dataset diversification. Future work includes integrating high resolution imagery, 3D structural analysis, and shading factors to enhance solar potential assessments.