Classification of Forest Stratification and Evaluation of Forest Stratification Changes over Two Periods Using UAV-LiDAR

Abstract

The demand for spatially explicit and comprehensive forest attribute data has continued to increase. Light detection and ranging (LiDAR) remote sensing, which can measure three-dimensional (3D) forest attributes, plays a significant role. However, only a few studies have used uncrewed aerial vehicle (UAV)-LiDAR to extract the characteristics of the 3D structure of the forest understory. Therefore, this study proposes a method for classifying and mapping forest stratification and evaluating forest stratification changes using multitemporal UAV-LiDAR data. The study area is a forest of approximately 25 ha on the west side of the Expo Commemorative Park (Suita City, Osaka Prefecture, Japan). Three-dimensional point cloud models from two measurement periods during the leaf-fall season were used. Forest stratification was classified using time-series clustering of 2024 data. The classification of forest stratification and its spatial distribution effectively reflected the actual site conditions. By applying time-series clustering, the forest stratification was successfully classified using only UAV-LiDAR data. Changes in forest stratification were evaluated using data from 2022 to 2024. In areas where changes in forest stratification were evaluated as significant, evidence of tree felling was confirmed. In addition, changes in forest stratification were quantitatively evaluated. The proposed method uses only UAV-LiDAR, which is highly versatile; thus, it is expected to apply to various forests. The results of this study are expected to deepen our ecological understanding of forests and contribute to forest monitoring and management.

1. Introduction

With the increasing demand for information about forest management, reporting, and research, there has been a continuous increase in the demand for spatially explicit and comprehensive forest attribute data [1]. Remote sensing is useful for obtaining forest attribute data because it is superior to conventional field surveys in terms of speed, coverage, and ability to describe forest attributes [2]. Furthermore, remote sensing can reliably implement spatially continuous forest monitoring, unlike information acquisition based on sampling plots [3].

Optical sensors, which are extensively used in remote sensing, can only acquire two-dimensional forest attributes, such as the horizontal distribution of vegetation, and cannot acquire three-dimensional (3D) forest attributes, such as forest stratification [4]. In contrast, light detection and ranging (LiDAR) remote sensing can acquire 3D forest structures because it can penetrate tree canopies [4,5]. Thus, LiDAR remote sensing can directly measure 3D forest structures, such as tree canopy heights, topographies under tree canopies, and vertical leaf distributions [6]. Forests have distinct vertical layers (overstory, understory, grass/shrub layers) that influence species composition, carbon cycling, and ecological processes [7,8,9]. The 3D forest attributes most commonly acquired by LiDAR remote sensing are tree height, diameter at breast height (DBH), and aboveground biomass [1]. Other important forest-monitoring attributes such as leaf area index (LAI) and canopy cover can also be acquired [2]. These 3D forest attributes are detected or estimated using models trained on LiDAR data [6,10]. In addition, in recent years, it has become possible to accurately detect or estimate 3D forest attributes without requiring parameters for each site using machine learning [3]. In addition, LiDAR remote sensing increases the potential to acquire 3D forest attributes in locations with difficult field access, such as mangrove forests [11]. Studies using LiDAR remote sensing have been conducted in various forest biomes around the world [12], and these studies have fundamentally changed how 3D forest structures are observed and described [2]. The very detailed 3D point cloud data provided by LiDAR remote sensing can revolutionize and expand our understanding of forest structures in forest ecosystems [2]. In addition, the 3D forest attributes acquired by LiDAR remote sensing support forest monitoring and decision making in forest management [4,13].

Depending on the platform used, LiDAR remote sensing can be divided into space-borne LiDAR scanning, airborne LiDAR scanning (ALS), uncrewed aerial vehicle (UAV) LiDAR scanning (ULS), and terrestrial LiDAR scanning (TLS) [14]. Of these, ULS has attracted significant attention in recent years because of its remarkable technological development [15]. By flying at a relatively low altitude and speed, ULS can obtain a point density that is several orders of magnitude higher than that of conventional ALS, and when combined with a wide scanning angle, ULS can resolve the structure of individual trees and branches, thereby achieving the same effect as TLS [14]. Therefore, ULS can facilitate the understanding of the characteristics of the 3D structure of the forest understory, which has not been possible until now [3,14]. The 3D structure of the forest understory is an important attribute related to vegetation diversity and wildlife habitats; however, its acquisition remains a challenge [16]. Jacon et al. [17] used ALS to classify the secondary succession stage based on differences in forest stratification. However, there is limited research on the extraction of the characteristics of the 3D structure of the forest understory using ULS. In addition, using multitemporal datasets can enable more accurate monitoring of changes in forests [18]. Multiperiod stratification enables the tracking of changes in forest cover and structure over time, supporting better understanding of forest dynamics and ecosystem services [19]. However, the effectiveness of multitemporal datasets in the analysis of 3D forest structures has not been clarified [20]. To address these deficiencies, in this study, ULS data were utilized to conduct stratified classification of the 3D structural characteristics of forest understory. Additionally, the two measurement periods’ ULS data were employed to investigate the changes in the forest stratification.

Therefore, this study aims to propose a new method for acquiring 3D forest attributes to support forest monitoring and decision making in forest management. To achieve this, I developed a method for classifying and mapping the forest stratification using only ULS data by applying time-series clustering, and a method for mapping changes in the forest stratification using ULS datasets collected at two periods.

2. Materials and Methods

2.1. Study Area

The study area was a forest of approximately 25 ha on the western side of the Expo Commemorative Park (Suita City, Osaka Prefecture, Japan) (Figure 1). The forest was artificially created on the site of the former World Expo, and approximately 50 years have passed since it was planted. Various trees were planted with reference to the potential vegetation in the area. At present, the growth of 73 species of trees has been confirmed, with evergreen broad-leaved trees accounting for 69% and deciduous broad-leaved trees accounting for 23%. Many of the trees are approximately 15 m tall. The forest is managed by the park’s administrators, who remove fallen trees, cut down trees along paths, and cut down trees for maintenance.

2.2. LiDAR Measurements

The DJI Matrice 300 was used with DJI L1 installed. DJI L1 had a repetitive scanning pattern and dual return settings. In this case, the horizontal and vertical fields of view were 70.4° and 4.5°, respectively; the scan rate was 480,000 points/s; and the system accuracy was 10 cm horizontally and 5 cm vertically. Using UgCS (SPH Engineering, Ķekavas novads, Latvia), a flight route was created with a flight altitude of 60 m and a side lap of 70%, and the created flight route was imported into DJI Pilot2 and measured at a flight speed of 8 m/s. During the measurement, a DJI D-RTK 2 mobile station was set up to correct the position using a real-time kinematic (RTK)-global navigation satellite system (GNSS). Data measured on 9 March 2022, and 28 February 2024, during the leaf-fall period were used.

2.3. Data Analysis

The measurement data were post-processed using DJI Terra to obtain 3D point cloud models (las). The 3D point cloud models (las) were imported into GreenValley LiDAR360 v8.0, and the data were cleaned using the UAV processing function. Although the measurements were taken using RTK-GNSS, a positional offset was observed between the 3D point cloud models from different periods. Therefore, the positional discrepancy was corrected using a 3D point cloud model obtained by processing images taken on 28 February 2024, using DJI P1 and photogrammetry. Photogrammetry was performed using Agisoft Metashape Pro. Ver. 2.1.3, and the default parameters were used. To improve the positional accuracy of the 3D point cloud model, 17 ground control points (GCPs) were set up at geographical features in the study area. The GCPs were measured using a GNSS receiver (Drogger DG-PRO1RWS, Biz Station Co., Ltd., Mathumoto, Japan) that can perform RTK-GNSS positioning. The positioning accuracy was ±10 mm or less horizontally and ±15 mm or less vertically. The 3D point cloud models generated by DJI L1 and DJI P1 were imported into PiX4D Survey. Using the registration function, the positions of the 3D point cloud models generated by DJI L1 were corrected based on the 3D point cloud model generated by DJI P1. The 3D point cloud models generated by DJI L1, whose position had been corrected, were re-imported into GreenValley LiDAR360 v8.0, and in order to improve the classification accuracy of point clouds of the ground surface, first the vegetation point clouds were classified using the classify-by-deep-learning function. Next, the point clouds of the ground surface were classified by the classify-ground-points-by-the-CSF function with the default parameters for gentle terrain. The point clouds of the ground surface were interpolated using the triangulated irregular network method to construct a 5 cm × 5 cm digital terrain model (DTM). The heights of the 3D point cloud model were corrected from elevation to ground height (normalized) using the DTM. The normalized 3D point cloud models were divided into 1 m intervals (0–18 m) and converted to geographic information system point data. The 5 m × 5 m grid set for the entire study area was used as the aggregation unit (n = 11,581), and the point data for each height were divided using the intersect function in Esri ArcGIS Pro. This data processing resulted in a dataset containing the number of points in the 3D point cloud model for each height in each grid (Figure 2). Although the data were cleaned using the UAV processing function, the point cloud density was high at locations where measurements overlapped. Therefore, for each grid, the ratio of the number of points in the 3D point cloud model at each height to the total number of points in the 3D point cloud model was calculated (

Table 1. The number of points in the 3D point cloud model and the basic statistical measures of the ratio of the number of points at each height.

		2022			2024
Height [m]	Total Number of Points	Average Value of Ratio of the Number of Points	SD of Ratio of the Number of Points	Total Number of Points	Average Value of Ratio of the Number of Points	SD of Ratio of the Number of Points
0–1	19,638,505	21.4	27.5	56,405,386	23.5	28.6
1–2	1,549,171	1.7	4.4	4,069,204	1.6	4.3
2–3	1,730,906	1.8	3.9	4,782,441	1.8	3.9
3–4	1,994,324	2.1	3.8	5,590,471	2.1	4.0
4–5	2,475,637	2.4	3.9	7,202,450	2.5	4.8
5–6	2,871,546	2.8	4.1	7,059,462	2.6	4.2
6–7	3,399,157	3.3	4.7	7,870,296	3.0	4.4
7–8	4,222,814	3.9	5.1	9,252,247	3.5	4.7
8–9	5,284,427	4.7	5.8	11,297,075	4.3	5.5
9–10	6,342,703	5.5	6.9	13,618,852	5.1	6.4
10–11	6,799,118	6.0	7.6	15,287,256	5.6	7.0
11–12	6,897,893	6.3	7.4	16,223,810	6.1	7.2
12–13	7,738,343	7.2	8.9	17,823,650	6.8	8.2
13–14	8,514,329	8.0	10.4	19,725,588	7.5	9.4
14–15	8,150,082	7.8	10.9	20,724,559	7.8	10.6
15–16	6,660,758	6.5	10.6	18,535,793	7.0	10.6
16–17	4,651,100	4.8	9.5	13,779,564	5.4	10.0
17–18	3,045,343	3.6	9.7	8,633,966	3.9	9.8

).

Using the data generated in this study, time-series clustering was used to classify forest stratification. Time-series clustering is a method for clustering the dynamic structure between time-series data using dissimilarity, and it is used in a wide range of fields, including economics, finance, medicine, ecology, environmental studies, and engineering [21]. In this study, time-series clustering was used to cluster the grid according to the ratio of point cloud data at each height using R version 4.4.2 and the TSclust package. The TSclust package integrates the general dissimilarity indices used in time-series clustering. Dynamic time warping was used as the dissimilarity measure, and the furthest neighbor method was used as the cluster distance. Time-series clustering was performed using the latest data from 28 February 2024.

The changes in forest stratification were evaluated by calculating the difference between the ratios of point clouds at different heights in each grid at two different times. The ratio on 9 March 2022 was subtracted from that on 28 February 2024 for each height. To evaluate the overall change, the absolute values of the differences were calculated and summed for each grid.

2.4. Verification

To verify the clustering results, a field survey of the stratification was conducted. The stratification based on the 3D point cloud model of each cluster was compared with that obtained from the field survey. Two to three survey points were set for each cluster (totaling 27 points, Figure 5). Quadrats measuring 10 m × 10 m were set up in the field, and the vegetation coverage (%) for each layer was visually estimated. Although the stratification based on the 3D point cloud model has a 1 m interval, it is difficult to estimate the stratification at a 1 m interval in the field. Therefore, referring to previous monitoring results at the study area, the layers were divided into 0–1 m, 1–6 m, 6–10 m, and 10 m– and surveyed. The ratio of the number of points in the 3D point cloud model was re-calculated for the same layer divisions to compare with the field survey results.

To verify the evaluation of changes in the stratification, the forest conditions were surveyed in the field. Thirty-two survey points were set at locations where significant changes in the stratification were evaluated (Figure 8). The survey points were identified using a map application in the field, and the forest conditions at those locations were confirmed. The survey was conducted on 25 February 2025.

3. Results

3.1. Classification of Forest Stratification

The dendrogram was cut at 180 and classified into 13 clusters (Figure 3). The pattern of the ratio of the number of 3D point clouds at each height for each cluster is shown in Figure 4. The spatial distribution of each cluster is shown in Figure 5. The results of the field survey are presented in Figure 6 and Figure 7.

C1 has a large point clouds ratios in the 0–1 m layer and is distributed along garden paths. On-site observations confirmed that there were many areas along garden paths and in areas with few trees. At two locations, the point clouds ratios in the 0–1 m layer was large. In the quadrats, no trees were observed in the 10 m– layer, but trees were visible in the surrounding quadrats.

C2, C3, C4, and C6 have large point clouds ratios in the 0–1 m and 6–10 m layers and were distributed as clustered forest patches. Clustered forest patches were also confirmed on site, with many areas exhibiting well-balanced stratification. While the ratios varied by location, trees were distributed across all layers. The point clouds ratios in the 1–6 m layer were small, but the vegetation cover was large in the field.

C7, C8, C9, and C10 have large point clouds ratios in the 10 m– and 0 m–1 m layers, but small point cloud ratios in the 1 m–6 m and 6 m–10 m layers, and were distributed as clustered forest patches in some areas. Clustered forest patches were also confirmed on site, but many areas were dominated by tall trees. The vegetation cover was large in the 10 m– layer, but small in the 1 m–6 m and 6 m–10 m layers in the field.

C5 has a small point cloud ratio in the 10 m– layer and was scattered. In the field, there were areas where trees were cut down and no trees were distributed in the 6 m–10 m and 10 m– layers and, conversely, areas where few trees were distributed in the 0 m–10 m layers.

C11 has a small point cloud ratio over all layers and was scattered. In the field, there were many areas with low tree density and only tall trees distributed.

C12 and C13 have an extremely small number of grids included in the cluster. They were not confirmed in the field.

3.2. Changes in Forest Stratification

The spatial distribution of changes in forest stratification is shown in Figure 8. In areas where the changes in forest stratification were evaluated as significant, evidence of shrub collection for transplanting, the felling of trees along paths, and the felling of trees for forest maintenance were confirmed (Figure 9). Areas with a cleared understory but a preserved overstory were also evaluated as experiencing significant forest stratification changes.

4. Discussion

4.1. Classification of Forest Stratification

The analysis of forest stratification using LIDAR remote sensing has been proposed by various methods, such as leaf area index [22], probability of gaps from the top of the canopy to a height z [6], and vertical complexity index [23]. However, these methods do not capture forest stratification. In contrast, Coops et al. [24] developed a model to estimate the vertical foliage distribution from ALS data. Jacon et al. [17] used ALS data to classify the secondary transition stage based on differences in forest stratification. Although these models capture forest stratification, they are constructed using sampling plot data. In this study, time-series clustering was used to successfully classify forest stratification using only ULS data. Although two types of local conditions were mixed in C5, C12, and C13 which could be considered clustering noise, the classification of forest stratification and the spatial distribution effectively reflected the local conditions. The forest stratification was well-balanced and distributed as a cluster of forest patches; thus, C2, C3, C4, and C6 were identified, and these clusters are important for conserving the biodiversity of forests in the study area. C5 can be divided into two types by increasing the cut level of the dendrogram. C5 is a location with a unique stratification due to forest management; such areas can also be separated as clusters. C12 and C13 contain an extremely small number of grids and are considered noise, but the pattern of the ratio of point cloud data for each cluster is unique, and such areas can also be separated as clusters. Cluster analysis is an exploratory method, and there is no universal method [25]. The results vary depending on the dissimilarity index used, the method of calculating cluster distance, and the cut-off level of the dendrogram. Further research is needed to explore appropriate clustering methods.

LiDAR data tend to be biased toward the upper layer [24]. In addition, LiDAR data acquired during the leaves-off season can accurately detect tree trunks within forests [26]. In this study, I was able to classify the stratification structure, including information on all layers, by using LiDAR data acquired during the leaves-off season. However, C2, C3, C4, and C6 have small point cloud ratios in the 1–6 m range, but the vegetation cover was large in the field. This suggests that the stratification is well-balanced, resulting in insufficient 3D point clouds in the 1–6 m range. The amount of information on the understory is affected by factors such as forest tree species composition, and further verification is needed. Forest stratification is an important forest attribute that supports forest monitoring and decision making in forest management [4,13]; however, obtaining forest stratification is a challenge [16]. The proposed method, which provides spatially explicit and comprehensive 3D forest attribute data, is expected to significantly contribute to forest monitoring and management.

4.2. Changes in Forest Stratification

Changes in forest structures over time are particularly important among forest attributes [1]. In this study, changes in the forest stratification structure were quantitatively evaluated by preparing data for classification at two different periods. Evidence of tree felling was found at sites where changes in forest stratification were assessed as significant; thus, the proposed method is considered effective for forest monitoring and management. In addition, areas with a cleared understory but a preserved overstory were assessed as experiencing significant forest stratification changes. This finding also demonstrates that the proposed method is innovative and can assess changes in forest structures that cannot be detected using optical sensors.

4.3. Potential Applications of Proposed Method

In deciduous forests, LiDAR data acquired during the leaf-fall period are suitable for understanding the 3D structure of the forest understory [3]. The study area is a mixed forest of deciduous and evergreen trees. Although evergreen broad-leaved trees are predominant, the forest stratification was understood using LiDAR data acquired during the leaf-fall period. The possibility of obtaining information below the tree canopy is affected by the probability of the laser penetrating the canopy; thus, this depends on the canopy structure [4]. Therefore, future studies should continue to verify the vegetation types to which the proposed method can be applied.

This study used a small-footprint discrete-return LiDAR sensor, which is commonly used in ULS; thus, it is considered highly versatile. By exploiting the advantages of low operating costs and immediacy of flight of UAVs [26], multitemporal data can be easily constructed, and changes in forest stratification over long periods can be monitored. In addition, the 3D structure of the forest understory is an important attribute related to vegetation diversity and wildlife habitat [16]; thus, it is hoped that the results of this study will be developed into analysis that combines the results with wildlife distribution.

5. Conclusions

This study successfully developed a method for classifying and mapping forest stratification and evaluating changes in the forest stratification using ULS multitemporal data. The proposed method uses only ULS data, which are highly versatile; thus, it is expected to apply to various forests. The potential of LiDAR remote sensing as a tool for deepening ecological understanding typically depends on the tools, methods, and approaches created by users [10]. The results of this study are expected to deepen our ecological understanding of forests and contribute to forest monitoring and management. The future challenges are to apply and verify the methods of this study in various forest types and to integrate the clustering results with habitat analysis of wildlife.