1. Introduction
European forests have recently faced an increasing number and severity of disturbance events, which are closely associated with climate change [
1]. Apart from abiotic factors such as wind or fire, insect pests and pathogens, both native and non-native, are on the rise. An example of such an unprecedented disturbance situation is the current bark-beetle outbreak in Central Europe [
2]. Important mechanisms for mitigating the effects of individual disturbance events are the prediction of the possible trajectory of their spread and the subsequent direction of controls or management interventions in these areas. Methods for quickly and reliably determining the territory’s condition are needed to involve such a strategy. The rapid development of remote sensing techniques represents a suitable tool for these investigations.
Light Detection And Ranging (LiDAR) is an active sensing technology that can provide a detailed 3D reconstruction of objects and scenes, such as forest trees and stands [
3,
4], and represents a practical way to measure tree dimensions [
5], volumes [
6], and other characteristics [
7]. LiDAR data’s characteristics and level of detail vary according to the carrier platform: highly detailed and accurate reconstructions are provided with stationary terrestrial laser scanning (TLS) [
8], while a higher efficiency of data acquisition is offered by mobile devices installed on terrestrial mobile platforms or carried by a human (mobile laser scanning, MLS) [
9] or piloted aircraft (aerial laser scanning, ALS). A special case of the latter method mentioned is data acquisition using unmanned aerial vehicles (UAVs) equipped with a LiDAR sensor—UAV laser scanning (ULS).
In recent years, UAV laser scanning has gained attention because of its safety, flexibility of data acquisition, and high-quality data compared to ALS, which is all caused by UAVs’ ability to fly at lower speeds and at lower altitudes above ground. The first study using a LiDAR-equipped UAV in forestry can be found as early as 2010 [
10], where the authors presented a novel system based on a combination of Ibeo Lux and a Sick LMS151 laser scanner system together with IMU and GPS units. The authors acquired 3D point clouds that enabled tree height measurement with an error of 30 cm. Later, Wallace [
11] presented a similar system additionally equipped with a high-definition (HD) camera for more precise point-cloud derivation. In the study, it was proven that the inclusion of the video information improved the accuracy of the final point cloud from 0.61 m to 0.34 m (root-mean-square error assessed against ground control).
Many scientific publications have dealt with the extraction of individual tree crowns and assessing dendrometric characteristics from ALS data [
12,
13,
14]. ULS data are characterized by markedly higher densities, reaching thousands of points per square meter, and a higher diversity of scan angles, therefore providing high-quality 3D representations of tree components, such as tree stems or branches. ULS data were successfully used to identify individual tree stems [
15,
16], reaching detection rates from 51% to 87% for trees with diameters over 15 cm and 50 cm, respectively [
17], and up to 91% [
18] or almost 100% [
19,
20], depending on the forest stand type and conditions. Studies evaluating the suitability of ULS data for stem diameter estimation from direct measurements in point clouds report that most of the stems can be measured with a root-mean-square error of 4 cm [
21] to 6 cm [
19] with data acquired using Riegl VUX-SYS. Contrarily, Hyppä et al. [
22], utilizing an identical scanning system for data acquisition based on diameter errors, concluded that the above-canopy UAV laser scanning method was insufficient for data collection at the individual tree level; however, under-canopy UAV laser scanning was able to provide data comparable with terrestrial methods [
23]. ULS data were also successfully utilized for calculating stand volume [
24].
It has been shown that LiDAR 3D data can be utilized to derive structural characteristics at the tree or stand level. Structural characteristics, or metrics, are an effective tool in many applications. Liu [
25], in their study, assessed the effectiveness of plot-level metrics (i.e., distributional, canopy volume, and Weibull-fitted metrics) and individual-tree-summarized metrics (i.e., maximum, minimum, and mean height of trees and the number of trees from the individual tree detection (ITD) results) derived from UAV–LiDAR point clouds. Then, these metrics were used to fit estimation models of six forest structural attributes within a Ginkgo plantation in east China using parametric (i.e., partial least squares (PLS)) and non-parametric (i.e., k-Nearest Neighbors (k-NN) and random forest (RF)) approaches. Another important application for tree metrics in LiDAR data is species classification [
26], or with the addition of spectral imagery e.g. Pereira Martins-Neto et al. 2023 [
27]. Holmgren and Persson [
28] tested the classification of Scots pine versus Norway spruce on an individual tree level using features extracted from airborne laser scanning data. Regarding ULS data, Krůček et al. [
17] applied a random forest classified on a set of geometric properties derived from individual tree point clouds and, with 85% accuracy, classified trees into broad-leaved and needle-leaved classes.
Nowadays, commonly used deep learning algorithms in forestry studies such as Support Vector Machine (SVM) and RF play an important role in remote sensing data classification. An overview of RF applications in forestry remote sensing data is presented by Belgiu et al. [
29]. SVM was described by Vapnik [
30], and Mountrakis et al. [
31] provide an overview of its usage in remote sensing. The high accuracy of SVM was highlighted by Krajnčíć [
32] in terms of vegetation classifications. These two deep learning algorithms were compared by Pal [
33], who proved that there are lower user-defined requirements for RF than for SVM.
In this study, we analyze the metrics of point clouds of individual trees that were obtained using the Riegl (RIEGEL Laser Measurement System GmbH, Horn, Austria) VUX-SYS system. The first reference to the use of these systems in the scientific literature is the work of Brede et al. [
21], where the authors describe in detail the system components and quality of derived Digital Terrain Models (DTMs), Digital Surface Models (DSMs), and Canopy Height Models (CHMs) from the resulting point clouds, which were compared with models that were based on point clouds of the same areas obtained with the TLS system RIEGL VZ-400 (RIEGEL Laser Measurement System GmbH, Horn, Austria).
In this study, we are looking for point-cloud metrics that are significantly different for broadleaf and coniferous trees using descriptive statistics and the above-mentioned deep learning algorithm RF. We assume that spruce trees have a regular “star” shape of the crown and that the distribution of branches has a regular structure. On the other hand, the crown of a broadleaf is expected to be asymmetric, and the branches are more frequent over the entire trunk. Therefore, we have chosen the Clark–Evans spatial aggregation index (CE), among others, to improve classification results using the metrics derived with descriptive statistics methods. CE has been, till now, used for the evaluation of tree distribution in forest stand but not for point clustering analyses on such data.
4. Discussion
The benefits of using the VUX-SYS were highlighted by Brede and Wiesser [
20,
21] with regard to the derivation of individual tree representative 3D point clouds along the whole stem profile, where they emphasized the possibility of canopy penetration and data representativeness mainly in top canopy parts. Both authors agree that ULS has the advantage of providing a relatively high canopy point density contrary to the TLS method. As this study shows, the significant predictors used for classification were located mostly in the upper part of the stem profile.
The differences between coniferous and broadleaf trees based on the differences in CHM with regard to seasonal canopy changes were described by Reitberger [
45]; however, this approach is also sensitive to deciduous coniferous trees because these tree species show similar canopy mean-height changes from the points belonging to the canopy as those of broadleaf tree species. Nevertheless, generating time series for canopy changes is a time-demanding method.
Tree canopy shape was also considered in this study. There were two main assumptions regarding the differences in the canopy. The first considers the bottom part, e.g., the beginning of the crown part along the stem profile, contrary to the second assumption regarding the top of the tree crown, where it was expected that the difference in shape [
46] could be explained according to the different numbers of points in the current percentile. The positive impact of the height normalization process for tree shape description was mentioned by Brandtberg [
47]. Regarding the treetop, the broadleaf trees were expected to have a smooth and round treetop contrary to coniferous trees with a sharper treetop. Both assumptions were confirmed with the derived GLM.
Leaf-on and leaf-off conditions were also compared by Reitberger, Liang, and Kim [
45,
48,
49], and because of the leaf-on conditions during the data acquisition for this study, the results may have been negatively influenced.
The accuracy of the RF was clearly influenced by the different study areas as mentioned by Vetrivel [
50]. In comparison to the accuracy of Pal [
33] (around 88%), in this study, we observed slightly higher accuracy (95%) but with five times (500) more RF decision trees, and we included the CE index. The study of Shen [
51] used a combination of airborne hyperspectral and LiDAR data for the classification of 18 tree species based on individual tree metrics, again reaching a similar accuracy but with a combination of two RS methods. A slightly lower accuracy (76%) was achieved by Pereira Martins-Neto et al. [
27] for eight tree species in tropical forests by UAV hyperspectral data and LiDAR metrics. Still, such forests' structural composition is more complex than European temperate or boreal forests. As Belgiu [
29] mentions, there have been numerous variable investigations aimed at prediction power; however, nowadays, the aim is per-pixel classifications. Nevertheless, RF is less sensitive to the training data [
29] than other machine learning algorithms, and this was the main reason for using it in this study.
RF classification based on geometric 3D shape and intensity was also used in conditions in the Czech Republic by Krzystek et al. [
52], who used ALS LiDAR data on a large scale-area for single tree detection and, afterward, dead tree detection. The LiDAR data were fused with multispectral data, and the overall accuracy reached over 90%. High accuracy in tree species classification from multitemporal Sentinel-2 data was attained by Immitzer et al. [
53], who highlighted multitemporal data for this data type.
Another classification method that should be used is SVM [
30] with regard to the recent trends in RS [
31,
32], where it is used for classifying two classes. However, the use of SVM and other deep learning algorithms for this kind of data classification needs more development.