Benchmarking individual tree segmentation using multispectral airborne laser scanning data: the FGI-EMIT dataset

TL;DR

FGI-EMIT introduces the first large-scale multispectral LiDAR ITS benchmark, enabling rigorous comparison between unsupervised and deep-learning approaches on 1,561 manually annotated boreal trees, including understory specimens. The study systematically benchmarks four unsupervised algorithms and four 3D DL models, finding that DL methods substantially outperform traditional geometry-based methods, with ForestFormer3D achieving the highest test F1-score of $73.3\%$ and markedly better understory detection than alternatives. An ablation reveals that multispectral reflectance generally does not improve DL-based ITS performance, though single-channel inputs can help some models, and MS information benefits may be model-specific. Across a density range down to $10$ points/m$^2$, DL methods remain superior, highlighting robustness of learned representations, while the dataset’s inclusion of built environment broadens ITS applicability. The authors publish FGI-EMIT to promote standardized benchmarking and future multispectral ITS research.

Abstract

Individual tree segmentation (ITS) from LiDAR point clouds is fundamental for applications such as forest inventory, carbon monitoring and biodiversity assessment. Traditionally, ITS has been achieved with unsupervised geometry-based algorithms, while more recent advances have shifted toward supervised deep learning (DL). In the past, progress in method development was hindered by the lack of large-scale benchmark datasets, and the availability of novel data formats, particularly multispectral (MS) LiDAR, remains limited to this day, despite evidence that MS reflectance can improve the accuracy of ITS. This study introduces FGI-EMIT, the first large-scale MS airborne laser scanning benchmark dataset for ITS. Captured at wavelengths 532, 905, and 1,550 nm, the dataset consists of 1,561 manually annotated trees, with a particular focus on small understory trees. Using FGI-EMIT, we comprehensively benchmarked four conventional unsupervised algorithms and four supervised DL approaches. Hyperparameters of unsupervised methods were optimized using a Bayesian approach, while DL models were trained from scratch. Among the unsupervised methods, Treeiso achieved the highest test set F1-score of 52.7%. The DL approaches performed significantly better overall, with the best model, ForestFormer3D, attaining an F1-score of 73.3%. The most significant difference was observed in understory trees, where ForestFormer3D exceeded Treeiso by 25.9 percentage points. An ablation study demonstrated that current DL-based approaches generally fail to leverage MS reflectance information when it is provided as additional input features, although single channel reflectance can improve accuracy marginally, especially for understory trees. A performance analysis across point densities further showed that DL methods consistently remain superior to unsupervised algorithms, even at densities as low as 10 points/m$^2$.

Figures (9)

• Figure 1: Overview of the study area in the Espoonlahti district of Espoo, Finland. (a) Map of Finland with the location of Espoonlahti highlighted. (b) Orthophoto of the Espoonlahti district from summer 2024 (30 cm pixel size). Image obtained from the espoo2024data. (c) Examples of test forest plots with manually generated instance annotations.
• Figure 2: Examples of original data and manually generated annotations for two forest plots (IDs 1001 and 1028). (a) Original point cloud with pseudo-colors generated from scaled reflectance values of scanners 1, 2, and 3 assigned to the red, green, and blue channels, respectively. (b) Instance annotations of individual trees, where each tree instance is shown in a distinct color and non-tree points are shown in gray. (c) Semantic annotations of the data, with each class assigned a distinct color.
• Figure 3: Visual examples of trees from each crown category. In (a)--(d), the tree belonging to the corresponding crown category is highlighted in red, while adjacent trees are shown in blue and non-tree points in gray. (a) Example of a tree from crown category A (plot 1013, tree number 12). (b) Example of a tree from crown category B (plot 1001, tree number 12). (c) Example of a tree from crown category C (plot 1022, tree number 168). (d) Example of a tree from crown category D (plot 1018, tree number 34).
• Figure 4: Visual comparison of instance predictions from the unsupervised individual tree segmentation algorithms. Predicted tree instances are shown in distinct colors, and non-tree points in gray. (a) Plot 1013. (b) Plot 1028. (c) Subsection of plot 1003. (d) Subsection of plot 1018.
• Figure 5: Visual comparison of instance predictions from the deep-learning-based individual tree segmentation models. Predicted tree instances are shown in distinct colors, and non-tree points in gray. (a) Plot 1013. (b) Plot 1028. (c) Subsection of plot 1003. (d) Subsection of plot 1018.