Loosdorf-MSL dataset: multispectral LiDAR data for LULC classification, supporting current and prospective European NMCAs' schemes

1. Overview

Loosdorf-MSL presents the first 3D multispectral (MS) LiDAR dataset for land use land cover (LULC) classification based on the current and prospective LULC classification schemes of European National Mapping and Cadastral Agencies (NMCAs).

By releasing Loosdorf-MSL, we aim to address the following gaps:

• The limited availability of publicly accessible MS LiDAR datasets for the development and evaluation of deep learning models.
• The need for LULC classification studies tailored to the practical requirements of NMCAs.
• The need to improve consistency and comparability across LULC products.
• Fine-grained 3D ground truth.

2. Dataset Characteristics

The Loosdorf-MSL dataset was acquired in October 2023 using a commercial MS airborne LiDAR system, RIEGL VQ-1560i-DW, operating at green (532 nm) and NIR (1064 nm) wavelengths over the village of Loosdorf and the city of Melk in Lower Austria. The Loosdorf-MSL dataset covers 1.7 km² of suburban and forested landscapes and comprises 103,238,318 manually labeled points categorized into eight and 20 classes, which are used for training and evaluating DL models. Table 1 reports the specifications of the Loosdorf-MSL dataset.

Photogrammetric point clouds and orthophotos (at 20 cm resolution) are generated by dense image matching using SURE nFrames and are provided in the Loosdorf-MSL dataset as auxiliary data to support future multimodal data integration studies. It is worth noting that the photogrammetric data do not cover the val4 plot.

For further information on Loosdorf-MSL data, please refer to our paper mentioned in Section 5. 

To check the LAZ files, LiDAR processing software like CloudCompare and OPALS can be used. We recommend QGIS software for checking orthophotos.

Table 1. Specifications of the Loosdorf-MSL dataset.

2.1 LiDAR Data Attributes

In addition to the spatial coordinates, the LiDAR point cloud includes eight attributes, summarized in Table 2, of which three are spectral. 

Table 2. LiDAR point cloud attributes.

3. Annotation Procedure and Labels Definition

Point clouds are manually annotated using CloudCompare software by three annotators. To ensure the accuracy of the annotated point clouds, the annotations are verified using auxiliary photogrammetric point clouds and field survey data. In addition, the annotation quality is further ensured through multiple rounds of independent visual inspection by the authors of the correlated paper to verify label consistency and minimize potential misannotations.

In the Loosdorf-MSL dataset, LULC classes are defined based on the current (GT_L1) and prospective (GT_L2) ALS-based LULC classification schemes of NMCAs. Table 3 provides descriptions of the semantic categories included in the Loosdorf-MSL dataset at both levels.

The Loosdorf-MSL dataset is divided into training, testing, and validation plots using a 71:17:12 ratio. This division maintains geographical independence among the training, validation, and testing sets as much as possible while ensuring that all classes at both levels of detail are represented in each subset whenever feasible. The dataset includes a total of 15 plots, of which six are used for training, four are used for testing, and the remaining plots are used for validation. 

Table 3. Description of the semantic labels available in the Loosdorf-MSL dataset.

4. What makes the Loosdorf-MSL dataset unique?

• The following characteristics make Loosdorf-MSL a unique dataset:

  • Multimodal data integration: Combines multispectral LiDAR data with multispectral aerial imagery.
  • Hierarchical annotations: Provides ground-truth annotations at two levels of detail.
  • Operationally relevant LULC classification: Annotations follow the NMCA's current (L1) and prospective (L2, fine-grained) LULC classification schemes.
  • High geometric accuracy: LiDAR and photogrammetric data are precisely georeferenced and co-registered through a hybrid adjustment procedure, ensuring spatial consistency between the two data sources.
  • Large spatial coverage: Covers approximately 1.7 km², enabling the development and evaluation of data-intensive deep learning methods.

A summary comparing the Loosdorf-MSL with other relevant benchmark datasets for LULC classification is presented in the Table 4. 

Table 4. Comparison of the Loosdorf-MSL benchmark dataset with other suburban/urban MS LiDAR benchmark datasets. MS LiDAR datasets are marked with *.

5. Citation

Any work using the data should cite the following paper:

Takhtkeshha, N., Rizaldy, A., Hollaus, M., Hyyppä, J., Remondino, F., Mandlburger, G., 2026. Loosdorf-MSL: Benchmarking deep learning models for European NMCAs’ LULC schemes with multispectral LiDAR. ISPRS Open Journal of Photogrammetry and Remote Sensing, 100154. https://doi.org/10.1016/j.ophoto.2026.100154.