Cross-Regional Urban Semantic Segmentation of Airborne LiDAR Point Clouds using Superpoint Transformers

Abstract

The increasing demand for accurate and scalable urban scene understanding has driven the development of deep learning methods for LiDAR point cloud semantic segmentation. This paper presents a cross-regional evaluation of a Superpoint Transformer (SPT) based segmentation framework within a unified processing pipeline. The system integrates airborne LiDAR datasets with varying characteristics, including DALES and Vaihingen, and performs preprocessing, feature extraction, and graph-based learning using superpoints. Unlike existing approaches that rely on domain adaptation, the proposed work focuses on analyzing generalization performance across different geographic regions without applying adaptation techniques. A shared label space is designed to enable consistent comparison between datasets. To the best of our knowledge, this is the first work to integrate the Vaihingen 3D ISPRS benchmark within the SPT framework and establish a cross-regional evaluation protocol between DALES and Vaihingen through a principled semantic class mapping. Experimental results demonstrate strong performance in in-domain scenarios, while a significant drop is observed in cross-domain evaluation due to variations in point density, urban structure, and data distribution. The study highlights the impact of domain shift and emphasizes the need for improved generalization in LiDAR semantic segmentation. The proposed framework provides a simple, efficient, and scalable approach for evaluating cross-regional performance in real-world applications.