EvoMS: multispectral LiDAR forest point clouds with partial semantic annotations

Description

This repository contains the multispectral LiDAR forest data set (EvoMS) associated with the paper Unsupervised deep learning for semantic segmentation of multispectral LiDAR forest point clouds (Ruoppa et al., 2025). EvoMS consists of 20 forest plots used to train GrowSP-ForMS, a leaf–wood separation model based on unsupervised deep learning, introduced in the same paper. The data set includes two manually annotated plots for evaluating model accuracy and 18 fully unlabeled plots. The point clouds have been normalized by subtracting the digital terrain model from the z-coordinates and ground points have been removed.

The multispectral point clouds in the data set were originally captured at the Scan Forest test site near Evo, Finland (61.19°N, 25.1°E) on 22nd of June 2021 using the Finnish Geospatial Research Institute’s (FGI’s) in-house developed laser scanning system, HeliALS-TW. For further details on data acquisition and preprocessing, please refer to the associated paper.

Labels

Forest plots with IDs 1 and 2 (plot_1.las and plot_2.las respectively) contain manually generated instance and semantic labels. All other plots are entirely unlabeled. Points belonging to the same tree instance have been manually assigned a common positive integer ID, stored in the tree_index field. Points not associated with any tree instance have tree_index set to 0. Semantic classes, which are stored in the classification field, are defined as follows:

Note that in the associated paper, points in the understory class were reassigned to the foliage class.  For more information on the manual labeling process, please refer to the paper.

Point cloud attributes

All point clouds in the EvoMS data set include the following non-standard attributes:

Code

The source code for GrowSP-ForMS, along with pretrained model weights and preprocessing scripts for the EvoMS data set, is available in this GitHub repository.

Data split

During preprocessing, the forest plots were divided into smaller, overlapping cylindrical point clouds (see the paper and source code), each of which was assigned a unique integer ID. The subset of cylinders from each labeled plot that was used as the test set is listed in the table below. All remaining data was used for training.

Citation

Any scientific publication using the data should cite the following paper:

Ruoppa, L., Oinonen, O., Taher, J., Lehtomäki, M., Takhtkeshha, N., Kukko, A., Kaartinen, H., and Hyyppä J., 2025. Unsupervised deep learning for semantic segmentation of multispectral LiDAR forest point clouds. ISPRS Journal of Photogrammetry and Remote Sensing, 228:694–722, 2025. doi:10.1016/j.isprsjprs.2025.07.038.

BibTeX:

@article{ruoppa2025unsupervised,
    author = {Lassi Ruoppa and Oona Oinonen and Josef Taher and Matti Lehtomäki and Narges Takhtkeshha and Antero Kukko and Harri Kaartinen and Juha Hyyppä}
    title = {{Unsupervised deep learning for semantic segmentation of multispectral LiDAR forest point clouds}},
    journal = {ISPRS Journal of Photogrammetry and Remote Sensing},
    volume = {228},
    pages = {694--722},
    year = {2025},
    doi = {10.1016/j.isprsjprs.2025.07.038},
}