NormalView: sensor-agnostic tree species classification from backpack and aerial lidar data using geometric projections

TL;DR

NormalView presents a sensor-agnostic projection-based framework that encodes local geometry as normal-vector colors in 2D projections to classify tree species from point clouds. Evaluated on high-density MLS backpack data and helicopter ALS data, it achieves strong MLS performance (OA around 95%) and competitive ALS performance when multispectral radiometry is used. The approach demonstrates that geometric projections, combined with strong CNN backbones like YOLOv11 and multi-view aggregation, can rival 3D methods, especially for minority species, and highlights the value of multispectral information. A publicly released MLS dataset supports reproducibility and further research in sensor-agnostic tree-species classification.

Abstract

Laser scanning has proven to be an invaluable tool in assessing the decomposition of forest environments. Mobile laser scanning (MLS) has shown to be highly promising for extremely accurate, tree level inventory. In this study, we present NormalView, a sensor-agnostic projection-based deep learning method for classifying tree species from point cloud data. NormalView embeds local geometric information into two-dimensional projections, in the form of normal vector estimates, and uses the projections as inputs to an image classification network, YOLOv11. In addition, we inspected the effect of multispectral radiometric intensity information on classification performance. We trained and tested our model on high-density MLS data (7 species, ~5000 pts/m^2), as well as high-density airborne laser scanning (ALS) data (9 species, >1000 pts/m^2). On the MLS data, NormalView achieves an overall accuracy (macro-average accuracy) of 95.5 % (94.8 %), and 91.8 % (79.1 %) on the ALS data. We found that having intensity information from multiple scanners provides benefits in tree species classification, and the best model on the multispectral ALS dataset was a model using intensity information from all three channels of the multispectral ALS. This study demonstrates that projection-based methods, when enhanced with geometric information and coupled with state-of-the-art image classification backbones, can achieve exceptional results. Crucially, these methods are sensor-agnostic, relying only on geometric information. Additionally, we publically release the MLS dataset used in the study.

Figures (11)

• Figure 1: Visualisation of the test site in Espoonlahti. (a): Location of the test site. (b): Orthophoto showing Espoonlahti. The tree segments, as well as the coverage of the different scanning methods, are visible. (c): A zoomed-in view of the orthophoto in (b), showing the tree segments in more detail. Figure adapted from taher2025.
• Figure 2: The same pine tree as captured by (a) the backpack mobile scanner, and (b) helicopter-ALS. The point counts of the segments are 399239.0 and 94586.0 respectively. The colouring of the points is purely for visualisation purposes.
• Figure 3: (a)-(b) Two examples of images used in training the models. The images are WOP images created from the same birch from the same angle. (a) All the points are used for the projection. (b) Points nearest to the "viewing plane" are removed to reduce occlusion on the trunk and inner branches. (c) Images in different resolutions of the same rowan. The resolution is achieved by bilinearly interpolating the image of size $1024\times1024$.
• Figure 4: Example images of the nine different tree species and image colouring methods used in model training. All samples are captured by HeliALS. For explanations for the colours, see section \ref{['sub:imagecreation']}.
• Figure 5: Species-wise metrics of the models trained on MLS data. (a) Species-wise precision, (b) species-wise recall, (c) species-wise $\text{F}_1$ scores. B&W refers to the model trained on black-and-white images.