Map2ImLas: Large-scale 2D-3D airborne dataset with map-based annotations

Airborne data are commonly used in mapping, urban planning, and environmental monitoring. However, deep learning (DL) for these tasks is often limited by the lack of large-scale multimodal datasets that represent diverse landscapes and detailed classes. In this work, we present Map2ImLas, a large-scale dataset created using topographic maps and high-resolution airborne data from the Netherlands. The dataset includes 2413 spatially matching tiles of maps, 2D orthoimages, digital surface models, and 3D point clouds, covering approximately 217 km2 across urban, suburban, industrial, rural, and forested areas. Map2ImLas provides per-pixel and per-point annotations for 20 different classes, applicable to both 2D and 3D semantic segmentation, as well as vector polygons for object delineation tasks. The proposed labeling process is fully automated, dynamically aligning all data sources to generate structured and consistent annotations. The pipeline is scalable and can be adapted to other regions in the Netherlands with minimal changes. We also introduce a DL-based workflow for labeling trees in 3D point clouds, using map data as semantic priors. To support DL applications, we provide a two-fold data split with non-overlapping training, validation, and test tiles. The dataset is benchmarked using several state-of-the-art 2D and 3D segmentation models to demonstrate its usability for semantic segmentation tasks. While the present evaluation focuses on segmentation, the structured vector annotations also enable future research on boundary extraction and object delineation. Overall, Map2ImLas reduces the need for manual annotation by reusing existing map data and supports geospatial AI in large-scale mapping and semantic labeling for multimodal data.