Multispectral airborne laser scanning for tree species classification: a benchmark of machine learning and deep learning algorithms

TL;DR

This study provides a comprehensive benchmark of machine learning and deep learning methods for tree species classification using multispectral airborne laser scanning data. It demonstrates that point-based deep learning, particularly Point Transformer models, outperform traditional ML and image-based DL on high-density multispectral ALS, achieving up to 87.9% overall accuracy and 74.5% macro-average on dense HeliALS data. Spectral information from multiple wavelengths substantially improves both overall and macro-average accuracy, and the authors reveal power-law scaling of classification error with training size, indicating faster gains for deep learning in large datasets. A crowdsourced field reference dataset and an international benchmarking framework enable robust cross-method comparisons and highlight practical considerations for deploying species classification at scale across different sensors and densities.

Abstract

Climate-smart and biodiversity-preserving forestry demands precise information on forest resources, extending to the individual tree level. Multispectral airborne laser scanning (ALS) has shown promise in automated point cloud processing and tree segmentation, but challenges remain in identifying rare tree species and leveraging deep learning techniques. This study addresses these gaps by conducting a comprehensive benchmark of machine learning and deep learning methods for tree species classification. For the study, we collected high-density multispectral ALS data (>1000 pts/m$^2$) at three wavelengths using the FGI-developed HeliALS system, complemented by existing Optech Titan data (35 pts/m$^2$), to evaluate the species classification accuracy of various algorithms in a test site located in Southern Finland. Based on 5261 test segments, our findings demonstrate that point-based deep learning methods, particularly a point transformer model, outperformed traditional machine learning and image-based deep learning approaches on high-density multispectral point clouds. For the high-density ALS dataset, a point transformer model provided the best performance reaching an overall (macro-average) accuracy of 87.9% (74.5%) with a training set of 1065 segments and 92.0% (85.1%) with 5000 training segments. The best image-based deep learning method, DetailView, reached an overall (macro-average) accuracy of 84.3% (63.9%), whereas a random forest (RF) classifier achieved an overall (macro-average) accuracy of 83.2% (61.3%). Importantly, the overall classification accuracy of the point transformer model on the HeliALS data increased from 73.0% with no spectral information to 84.7% with single-channel reflectance, and to 87.9% with spectral information of all the three channels.

Figures (16)

• Figure 1: (a) The Espoonlahti test site is located in Southern Finland. (b) Orthophoto of the Espoonlahti test site together with collected tree segments colored by the species in the reference dataset. The coverage of the HeliALS data is shown with the light blue line. (c) A close-up of the orthophoto shown in (b).
• Figure 2: Photograph of the scanner arrangement of the HeliALS system, showing the VUX-1HA scanner at $\lambda=1550$ nm, miniVUX-1DL scanner at $\lambda=905$ nm, and VQ-840-G scanner at $\lambda=532$ nm.
• Figure 3: (a) Example multispectral point clouds and photographs of the nine different tree species found in the dataset. The red, green and blue color channels of the point clouds represent the laser pulse return intensities at the wavelengths $\lambda_1 = \text{532 nm}, \lambda_2 = \text{905 nm}$ and $\lambda_3 = \text{1550 nm}$, respectively. The height scale varies between the visualized point clouds. Photographs of the representative species instances were taken during the autumn of 2024. (b) Number of segments per species in the training and test sets. (c) Species-wise tree height distributions for the entire dataset.
• Figure 4: Screenshot of the crowdsourcing application used to collect ground-truth species data. In this example, the application shows a true orthophoto map, segment boundaries (red lines), the surveyor's location (blue circle) based on GNSS positioning, and the information of the nearest tree together with possible actions for the surveyor. The true orthophoto map in the background was obtained from the espoo_open_data.
• Figure 5: (a) Point cloud projections of segment instances belonging to the different profile categories. (b) Profile category distribution by species for the test set.