Loosdorf-MSL: Benchmarking deep learning models for European NMCAs’ LULC schemes with multispectral LiDAR

Land Use Land Cover (LULC) classification is fundamental to national 3D mapping programs, geospatial analyses, resource management, and sustainable development planning. While multispectral (MS) LiDAR (Light Detection and Ranging) technology offers synchronized spatial-spectral information and Deep Learning (DL) has shown promise in 3D semantic segmentation, several challenges remain. Specifically, the application of 3D DL models to MS LiDAR data has received limited attention. Moreover, consistent class definition schemes aligned with the requirements of National Mapping and Cadastral Agencies (NMCAs), along with publicly available MS LiDAR benchmark datasets based on such schemes are still lacking. Furthermore, the potential of MS LiDAR data for supporting NMCA applications has not been systematically evaluated, limiting its integration into operational mapping workflows. Therefore, this study addresses these gaps by reporting NMCAs’ current (L1) and prospective (L2) LULC classification schemes, presenting a novel LiDAR benchmark dataset at two levels of detail (L1 and L2), evaluating seven state-of-the-art DL models, and finally conducting rigorous spectral ablation analyses at the two levels of detail. Benchmarking results demonstrate the superiority of the Point Transformer V3 model at both levels, achieving high mean Intersection over Union (mIoU) values of 78.4% for LULC-L1 (eight classes) and 58.9% (20 classes) for LULC-L2 by leveraging a two-wavelength commercial MS LiDAR system (532 nm and 1064 nm). Furthermore, spectral ablation studies highlight the significant potential of MS LiDAR data over purely geometric features. While the spectral advantage is marginal at LULC-L1 (a 1.1 Percentage Points (PP) improvement in mIoU by leveraging a vegetation index), it becomes considerably more pronounced at LULC-L2 (a 7.8 PP improvement). Our findings underscore the capacity of LiDAR reflectance and MS LiDAR for surface material identification, supporting the prospective transition of NMCAs schemes toward fine-grained LULC classification. The Loosdorf-MSL dataset contributes to improving consistency across LULC products. The dataset and code are available at https://github.com/3DOM-FBK/3D-LULC-classification .