DEEP LEARNING FOR SEMANTIC 3D CITY MODEL EXTENSION: MODELING ROOF SUPERSTRUCTURES USING AERIAL IMAGES FOR SOLAR POTENTIAL ANALYSIS

Abstract

On a global scale, semantic 3D city models with Level of Detail 2 become more and more available. Automated generation of higher Level of Detail models is an active field of research, but low coverage of dense LiDAR or photogrammetric point clouds is a barrier. Therefore, this paper presents a novel approach for enriching semantic 3D city models with roof superstructures extracted from aerial images using deep learning. The method maps and classifies superstructures in 2D and subsequently transforms them to 3D. Furthermore, we examine the benefit of the enriched model for solar potential analysis. The accuracy of solar potential analysis is improved by avoiding invalid simplifications of slope, shadow and panel placement. The enriched model reduces overestimation of accumulated solar potential by around 20% compared to an analysis based on aerial images only. The novel method contributes to increasing the availability of Level of Detail 3 models for larger areas, while posing further research opportunities.