Advancing Semantic Segmentation for Building Detection in Very High-Resolution Data

Abstract

Accurate building footprints are essential for urban planning, crisis management, and social science research, yet Germany lacks a comprehensive and up to date nationwide register. Existing sources such as cadastral data, and OpenStreetMap remain incomplete or inconsistent. At the same time current deep learning models for automatic footprint extraction still suffer from systematic errors. This thesis investigates whether optimizing preprocessing, postprocessing, and training data selection can improve the CNN-based extraction model proposed by Stiller et al. [2023]. Experiments with normalization strategies, LiDAR-based height layers, tile overlap, and threshold settings show that targeted adjustments enhance performance. The final pipeline, using DSM data and refined data normalization, improved Overall Accuracy by 7.0 percentage points, IoU by 5.2 percentage points, and F1 score by 3.3 percentage points compared to the baseline. A case study on Berlin illustrates the practical value of the generated data for the social sciences by linking building geometries with demographic and building use data. The findings highlight both the technical and applied relevance of the improved workflow: advancing footprint extraction toward official usability while enabling new insights in the social sciences.