Structural Changes in Boreal Forests Can Be Quantified Using Terrestrial Laser Scanning

: Terrestrial laser scanning (TLS) has been adopted as a feasible technique to digitize trees and forest stands, providing accurate information on tree and forest structural attributes. However, there is limited understanding on how a variety of forest structural changes can be quantified using TLS in boreal forest conditions. In this study, we assessed the accuracy and feasibility of TLS in quantifying changes in the structure of boreal forests. We collected TLS data and field reference from 37 sample plots in 2014 (T1) and 2019 (T2). Tree stems typically have planar, vertical, and cylindrical characteristics in a point cloud, and thus we applied surface normal filtering, point cloud clustering, and RANSAC-cylinder filtering to identify these geometries and to characterize trees and forest stands at both time points. The results strengthened the existing knowledge that TLS has the capacity to characterize trees and forest stands in space and showed that TLS could characterize structural changes in time in boreal forest conditions. Root-mean-square-errors (RMSEs) in the estimates for changes in the tree attributes were 0.99-1.22 cm for diameter at breast height (Δdbh), 44.14-55.49 cm 2 for basal area (Δg), and 1.91 - 4.85 m for tree height (Δh). In general, tree attributes were estimated more accurately for Scots pine trees, followed by Norway spruce and broadleaved trees. At the forest stand level, an RMSE of 0.60-1.13 cm was recorded for changes in basal area-weighted mean diameter (ΔD g ), 0.81-2.26 m for changes in basal area- weighted mean height (ΔH g ), 1.40-2.34 m 2 /ha for changes in mean basal area (ΔG), and 74 -193 n/ha for changes in the number of trees per hecta re (ΔTPH). The plot -level accuracy was higher in Scots pine-dominated sample plots than in Norway spruce-dominated and mixed-species sample plots. TLS-derived tree and forest structural attributes at time points T1 and T2 differed significantly from each other ( p < 0.05). If there was an increase or decrease in dbh, g, h, height of the crown base, crown ratio, D g , H g , or G recorded in the field, a similar outcome was achieved by using TLS. Our results provided new information on the feasibility of TLS for the purposes of forest ecosystem growth monitoring.


Introduction
Forests change over time and across space due to natural phenomena and anthropogenic processes. Biotic changes, such as forest growth and damage, affect forest structure and tree stems and branches grow annually in width and height. Conifer trees are typically evergreen but leaves of deciduous trees emerge in the spring and fall in the autumn. Forest growth process is also linked to the decomposition process which affects the amount of dead wood [1]. Insect damage, the spread of pathogens, wind damage, and forest fires, shape forests at different spatial and temporal scales. Insect damage, that is often combined with the spread of pathogenic fungi, causes defoliation and changes in stem forms. Of the 35 investigated trees, which were mainly conifers, changes in stem taper, cylindrical form factor, form quotient, and stem slenderness were analysed. The changes in the stem taper varied from -34% to 9%, the cylindrical form factor from 1% to 18%, the form quotient from 4% to 35%, and the stem slenderness from -2% to 6%.
The objective of this study was to strengthen the understanding how TLS can be used to capture boreal forest structure in space and time. Concentrating on the change of only a few attributes with a relatively low number of samples has been the common denominator for all the previous studies related to the use of TLS in quantifying changes in tree attributes. A more comprehensive investigation on the performance of capturing changes in tree and forest structural attributes with a large number of trees and sample plots is needed to assess the feasibility of TLS in monitoring forest structural changes. To fill in the knowledge gap, we used bi-temporal TLS data and field inventory data to cover changes in the structure of 1280 trees and 37 sample plots from varying boreal forest conditions. We hypothesized that TLS has the capacity to accurately estimate tree and forest structural attributes at single time points, and the estimated attributes at the beginning of a monitoring period significantly differ from the respective attributes at the end of the monitoring period. We assessed the performance of change quantification by using the most common tree and forest structural attributes. At tree level, we analysed changes in dbh (Δdbh), basal area (Δg), tree height (Δh), diameter-height ratio (Δd-h-ratio), height of the crown base (Δhc), and crown ratio (Δcr). At stand (sample plot) level, we analysed changes in basal-area weighted mean diameter (ΔDg) and -height (ΔHg), mean basal area (ΔG), and number of trees per hectare (ΔTPH). Thus, we used a range of attributes describing changes in horizontal and vertical forest structure.

Study materials
The study materials consisted of a multi-scan TLS data and a field inventory data acquired in 2014 (T1) and 2019 (T2) covering 1280 trees in 37 sample plots. The study site is located in Evo, southern Finland (61°19.6'N 25°10.8'E) where 91 sample plots (32 m x 32 m) were initially established in 2014 to cover the structural variation of forests (see e.g. [36]). A TLS data-acquisition campaign was carried out in spring 2014 using a Leica HDS6100 (Leica Geosystems, St. Gallen, Switzerland) and a Faro Focus 3D X330 (Faro Technologies Inc., Lake Mary, FL, USA) phase shift scanners, both operating at 1550 nm wavelength and measuring 508,000 points per second, delivering a hemispherical (310° vertical x 360° horizontal) point cloud with an angular resolution of 0.018° in both vertical and horizontal direction. A multi-scan approach was used to obtain a comprehensive point cloud for each sample plot by merging point clouds from five separate scanning locations. The scan setup consisted of one centre scan located at the plot centre and four auxiliary scans at quadrant directions (i.e. northeast, southeast, southwest, and northwest) about 11 m away from the plot centre. Artificial reference targets were used to register the point clouds together. Trees from each sample plot were located by manually detecting stem cross sections from horizontal TLS point cloud slices to construct a stem map. The stem maps were verified in the field and completed with the locations of small undergrowth trees that were not visible in the point cloud. A tree-wise field inventory was carried out in the summer of 2014 to acquire reference measurements of the tree attributes for T1. Tree species, dbh, h, hc, and health status (alive/dead) were recorded for all the trees with dbh exceeding 5 cm. Tree attributes were then aggregated at the plot level to obtain forest structural attributes: Dg (cm) and Hg (m), G (m 2 /ha), and TPH (n/ha). More detailed description on the TLS data acquisition and field inventory for T1 can be found in [22].
Field inventory and TLS data-acquisition campaigns were repeated for 37 sample plots in autumn 2019 to capture a five-year growth period in between the measurements. The stem maps were updated in the field with missing trees (i.e. trees fallen or harvested during the time period) and in growth of trees (i.e. trees with dbh exceeding the 5 cm threshold during the time period) while dbh and tree height were re-measured for all trees, and height of the crown base for Scots pine trees (Pinus sylvestris L.). TLS data at T2 was collected using a Leica RTC360 3D (Leica Geosystems, St. Gallen, Switzerland) time-of-flight scanner that operates at 1550 nm wavelength and measures 2,000,000 points per second, delivering a hemispherical (300° vertical x 360° horizontal) point cloud with an angular resolution of 0.009° in both vertical and horizontal direction. Similar to T1 TLS data acquisition, a multi-scan approach (i.e. one centre scan with four auxiliary scans) was used in the T2 TLS campaign to ensure consistent point cloud quality. The scan setup for T2 was slightly modified from the one used in the T1 campaign based on experience gained in [10] that suggests placing the auxiliary scans approximately at the plot borders to improve point cloud completeness (see Figure  1). Artificial reference targets and a Leica Cyclone 3D Point Cloud Processing Software were used to register the separate point clouds together for each sample plot.
Later on, we found that, due to the slightly different scan setups used in the TLS-campaigns, the overall quality of the point clouds was somewhat poorer in T1 because the auxiliary scans in T1 were located closer to the plot center than in T2 (see Figure 1). Therefore the sample plots were reduced in size, from the rectangular 32 m x 32 m plots (1024.0 m 2 ) to circular sample plots with a 11-m radius (380.1 m 2 ), to achieve more comparable point clouds between time points T1 and T2 (i.e., to ensure that most of the trees were scanned from multiple directions). Thus, the total number of fieldmeasured trees was 1280, of which 270 (21.1%) were Scots pine trees, 649 (50.7%) were Norway spruces (Picea abies (L.) H. Karst.), and 361 (28.2%) were broadleaved trees, mainly birches (Betula sp.) and European aspen (Populus tremula L.). The main tree species was defined for the sample plots based on proportional field-measured G at time point T2 by tree species. A sample plot was classified as a single species-dominated sample plot if any of the tree species accounted for more than 67% of the total basal area. Respectively, a sample plot was classified as a mixed-species sample plot if G of two or more tree species each accounted for more than 30% of the total basal area. This classification resulted in three groups of sample plots as 9 sample plots were classified as Scots pine-dominated, 13 sample plots were classified as Norway spruce-dominated, while 15 sample plots were classified as mixed-species sample plots. The sample plots covered a wide range of forest structural variation (see Table 1). Table 1. Variation in tree attributes by tree species and forest structural attributes by main tree species of a sample plot, measured in 2014 (T1) and 2019 (T2). Diameter at breast height (dbh), basal area (g), tree height (h), diameter-height-ratio (d-h-ratio), height of the crown base (hc), and crown ratio (cr) were defined at tree level whereas basal area-weighted mean diameter (Dg) and -height (Hg), mean basal area (G), and number of trees per hectare (TPH) were defined at plot level. Scots pine-dominated plots = basal area of Scots pine (Pinus sylvestris L.) accounts for more than 67% of the total basal area of a plot. Norway spruce-dominated plots = basal area of Norway spruce (Picea abies (L.) H. Karst. ) trees accounts for more than 67% of the total basal area of a plot. Mixed-species plots = G of two or more tree species each account for more than 30% of the total basal area.

Deriving tree and forest structural attributes from point clouds
A point cloud processing method presented in [37] was used in this study to measure tree attributes and to estimate forest structural attributes from the multi-scan TLS point clouds for each sample plot and both T1 and T2 (see Figure 1). First, a raster-based canopy segmentation was carried out to partition the point clouds into smaller areas using a canopy height model (CHM) at a 20 cm resolution. Variable Window Filter approach [38] was used to identify the local maxima in the CHM, and Marker-Controlled Watershed Segmentation [39] was applied to delineate canopy segments. The point cloud was then split according to the extracted crown segments using point-in-polygon approach. The crown-segmented point cloud was further classified into stem and non-stem points based on a general assumption that stem points had more planar, vertical, and cylindrical characteristics than points representing branches and foliage [10,15]. These characteristics were Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 8 July 2020 doi:10.20944/preprints202007.0154.v1 distinguished by applying surface normal filtering, point cloud clustering, and Random Sample Consensus (RANSAC)-cylinder filtering on horizontal point cloud slices. A more detailed description of the point cloud classification procedure can be found in [37].

Figure 1.
Outline of the TLS data processing workflow. One center scan and four auxiliary scans were used to acquire a multi-scan TLS point cloud data in 2014 and 2019 (a). TLS-based canopy height model (CHM) and a Marker-Controlled Watershed Segmentation procedure was applied to TLS point clouds to detect individual trees. Surface normal filtering, point cloud clustering and Random Sample Consensus (RANSAC)-based cylinder filtering were applied to identify vertical, cylindrical, and planar surfaces to classify TLS point cloud into stem and non-stem points (for more details, see [37]) (b). Then, the classified point cloud was used to extract tree attributes, namely diameter-at-breastheight (dbh) and tree height (h) for all the trees, and height of the crown base (hc) and crown ratio (cr) for Scots pine (Pinus sylvestris L.) trees (c).
Tree attributes, namely dbh, g, h, d-h-ratio, hc, and cr were extracted from the classified point cloud (see Figure 1, Table 2). Dbh and h were measured following the procedure originally presented in [10]. In other words, h was determined as the vertical distance between the highest and lowest points for each tree. Stem taper curve was estimated to determine dbh by measuring diameters through circle fitting at 20-cm vertical intervals to the stem points. The outliers in diameter-heightobservations were filtered out by comparing the measured diameters to the mean of three previous (or three closest at the bottom of the stem) diameters. Then a cubic spline curve was fitted to the diameter-height -observations to level unevenness in diameter measurements and to interpolate the missing diameters as suggested in [40]. Dbh was then obtained as the diameter at 1.3 m height from the taper curve. Tree-level g was computed from dbh measurements by considering the stem cross section as a circle (g = π * dbh 2 / 4) and d-h-ratio was computed as a ratio between dbh and h (d-hratio = dbh / h). Hc was determined by first binning the non-stem points into horizontal slices with a height of 20 cm, then computing a convex hull around the bin points projected to XY-plane and hc was determined at the height where the convex hull area exceeded a 1.5 m 2 threshold and the perimeter-to-area ratio for the convex hull was smaller than 2 ( Figure 1c). Cr was computed as the proportion of the height of a living crown from the tree height (cr = (h -hc) / h). Finally, the forest structural attributes (i.e. Dg, Hg, G, and TPH) were computed by aggregating the tree-level attributes at the plot level (Table 2). Dg and Hg were computed as a basal area-weighted mean from the TLSderived dbh and h measurements whereas G was a sum of g and TPH was the total number of trees within a sample plot per unit area. Table 2. Description of the attributes used in this study to characterize trees and forest structure in 2014 (T1) and 2019 (T2). A = sample plot area in hectares, n = number of trees within a sample plot.

Assessing the performance of the TLS-based method in quantifying changes in forest structure
Performance of the TLS-based method to quantify changes in tree attributes and forest structure was assessed by comparing the TLS point cloud-derived tree and forest structural attributes with the field-measured counterparts. For each TLS-derived tree, a corresponding field-measured tree was searched based on its spatial location. Capability of the TLS-based method to detect trees from the point clouds was then assessed by using completeness as an accuracy measure, indicating how large a part of the trees was detected from the point clouds. Accuracy of the TLS point cloud-derived estimates for tree and forest structural attributes at time points T1 and T2 as well as their difference (Δ) was assessed by using bias (mean error) and root-mean-square-error (RMSE) as accuracy measures: where n is the number of trees or sample plots, ̂i is the TLS point cloud-derived tree attribute or forest structural attribute for tree or plot i, and Xi is the corresponding attribute based on reference measurements. Relative RMSE (RMSE%) and bias (bias%) were computed by dividing the absolute RMSE and bias with average value of the respective field-measured attribute. For tree attributes, the accuracy was assessed by tree species (Scots pine, Norway spruce, and broadleaved trees) while for forest structural attributes the accuracy was assessed by main tree species of a sample plot (Scots pine-dominated, Norway spruce-dominated, and mixed-species sample plots). Coefficient of determination (R 2 ) was used to measure the relationship between the TLS point cloud-derived and field-measured tree and forest structural attributes. Paired-sample T-tests were used to test whether the TLS-based estimates for tree and forest structural attributes at time point T1 significantly differed from the respective estimates at time point T2.

Performance of detecting trees using bi-temporal TLS data
Out of the total number of 1280 trees that were measured in the field, 795 trees (62.1%) were detected from the point clouds at both timepoints T1 and T2. The detected trees represented 84.5% of the total basal area of all trees. Tree detection accuracy was highest among Scots pine trees, followed by Norway spruce and broadleaved trees ( Table 3). Trees that remained undetected were mainly small in size, while large trees were detected at both timepoints with high accuracy (Figure 2). Table 3. The number of trees by tree species that were measured in the field (Nref) and detected from the TLS point clouds (NTLS) at the beginning (2014) and at the end of the monitoring period (2019).
Completeness indicates how large a part of the detected trees represented the total number and the total basal area of the field measured trees.

Performance of characterizing trees in space using TLS
Strong relationships (R 2 = 0.99) between the field-measured and TLS-derived estimates for dbh and g were recorded at both timepoints (Figure 3a-b, 3d-e). Dbh was estimated with an RMSE of 1.18 cm (5.7%) and 0.90 cm (4.1%) at T1 and T2, respectively, with no significant (p > 0.05) differences in accuracy between the tree species (Table 4). However, g was estimated more accurately for Norway spruce (RMSE% 6.2-10.6%) than for Scots pine (RMSE% 12.5-13.0%) and broadleaved trees (RMSE% 14.5-16.0%). On average, dbh of Norway spruce were overestimated by 0.26-0.30 cm, while dbh of Scots pine and broadleaved trees were underestimated by 0.43-0.32 cm and 0.03-0.19 cm, respectively. Considering all the trees, slightly lower RMSEs and biases of dbh and g were obtained at T2 than at T1.
Relationship between the field-measured and TLS-derived estimates for h was stronger for Scots pine and Norway spruce (R 2 = 0.84-0.92) than for broadleaved trees (R 2 = 0.31-0.39; see Figure 3g-h). In general, TLS-derived h was underestimated, and the accuracy of h estimates were improved from T1 to T2 (Table 4). For Scots pine trees, RMSEs of 2.52 m (13.7%) and 2.42 m (12.1%) were recorded at time points T1 and T2, respectively. The estimation accuracy for Norway spruce was slightly lower being 3.99 m (20.1%) for T1 and 3.03 m (14.4%) for T2. For broadleaved trees, the estimation accuracy decreased even more, being 6.26 m (31.7%) at T1 and 6.56 m (30.8%) at T2. Also, differences in the accuracy of h estimates led to differences in the accuracy of d-h-ratio estimates between the tree Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 8 July 2020 doi:10.20944/preprints202007.0154.v1 species. R 2 of 0.76-0.79 was recorded between the field-measured and TLS-derived d-h-ratio for Scots pine, while the respective values for Norway spruce and broadleaved trees were 0.58-0.73 and 0.43-0.51, respectively (Figure 3j-k). Hc and cr were estimated for Scots pine trees only. R 2 between the field-measured and TLSderived estimates for hc and cr were 0.78-0.79 and 0.52-0.58, respectively (Figure 3p-q). For hc estimates an RMSE of 2.71 m (25.7%) for T1 and 2.55 m (20.4%) for T2 with an underestimate of 1.36-1.70 m (Table 4). On average, cr was overestimated by 10.3-11.7% with a relative RMSE of 22.5-22.8% (Table 4).

Figure 3.
Relationship between the field-measured reference measurements and the TLS-derived estimates on tree attributes, such as diameter-at-breast-height (dbh), basal area (g), tree height (h), diameter-height-ratio (d-h-ratio), and crown ratio (cr) by tree species measured in 2014 (T1) and 2019 (T2) as well as their difference (Δdbh, Δg, Δh, Δd-h-ratio, and Δcr). The solid black line represents the 1:1 relationship between the reference and the estimated values. R 2 p, R 2 s, and R 2 b denote coefficient of determination for Scots pine, Norway spruce, and broadleaved trees, respectively. Table 4. Bias and root-mean-square-error (RMSE) of TLS-derived estimates for tree attributes, namely diameter-at-breast-height (dbh), basal area (g), tree height (h), diameter-height-ratio (d-h-ratio), height of the crown base (hc) and crown ratio (cr) by tree species at time points T1 (2014) and T2 (2019) as well as their change (Δ). Negative bias denotes underestimation.

Performance of characterizing tree attributes in time using TLS
It was possible to quantify changes in tree attributes using TLS. Paired sample T-tests showed that the tree attributes estimated at T1 significantly (p < 0.01) differed from the respective attributes estimated at T2 (Table 5). Relationship between the changes in field-measured and TLS-derived tree attributes was stronger for attributes characterizing changes in horizontal tree structure (i.e., Δdbh and Δg) than attributes characterizing changes in vertical tree structure (i.e., Δh, Δhc, and Δcr; Figure  3). In general, the changes in tree attributes were estimated with smaller RMSE for Scots pine than for Norway spruce and broadleaved trees (Table 4). RMSE in Δdbh estimates was 0.99-1.22 cm (83.9-103.5%), and most often Δdbh was overestimated by 0.04-0.17 cm (3.8-15.0%) depending on tree species. Differences in the accuracy of Δg estimates between the tree species were small and not considered statistically significant (p > 0.05). On average, Δh was overestimated by 0.21-0.90 m (13.8-72.3%) with an RMSE of 1.91-4.85 m (122.8-321.5%). In this case, the highest accuracy was recorded for Scots pine, followed by Norway spruce and broadleaved trees. The accuracy of Δd-h-ratio was considerably lower than the accuracy of other tree attributes. Relative errors in Δd-h-ratio were large (RMSE 942.6-4516.2%) due to uncertainty in tree height estimates.
Attributes characterizing changes in crown structure were estimated for Scots pine trees only. With TLS, 34% and 35% of the variation in Δhc and Δcr respectively, was possible to capture. On average, Δhc was overestimated by 0.34 m (20.7%) with an RMSE of 1.97 m (120.6%) which is at the same level to the accuracy of Δh estimates for Scots pine trees (Table 4). Relative errors in Δcr estimates were larger (RMSE% 318.1%) due to uncertainty in h and hc estimates. Table 5. The p-values from the paired-sample T-tests indicating the significance of the differences between the TLS-derived estimates of tree attributes, such as diameter-at-breast-height (dbh), basal area (g), tree height (h), diameter-height-ratio (d-h-ratio) and crown ratio (cr) by tree species measured in 2014 (T1) and 2019 (T2).

Performance of characterizing forest structural attributes in space with TLS
Forest structural attributes were estimated accurately with TLS point clouds from the both timepoints. Strong relationships (R 2 > 0.86) between the field-measured and the TLS-derived estimates for Dg, Hg, G, and TPH were recorded for Norway spruce-dominated sample plots ( Figure  4). On Scots pine-dominated sample plots, the R 2 indicated stronger relationship between the fieldmeasured and the point cloud-derived estimates for Dg and Hg (R 2 > 0.96) than for TPH (R 2 = 0.46-0.66) and G (R 2 = 0.21-0.34). On mixed-species sample plots, an R 2 of 0.93-0.95 was recorded for Dg, 0.50-0.64 for Hg, 0.63-0.72 for G, and 0.44-0.63 for TPH. Considering all the sample plots, the estimation accuracy for Dg and G was at the same level for T1 and T2 while Hg and TPH estimates seemed to be slightly more accurate at T2 (RMSE% 7.9% for Hg and 49.3% for TPH) than at T1 (RMSE% 11.2% for Hg and 58.5% for TPH; Table 6).

Performance of characterizing forest structural attributes in time using TLS
Changes in forest structural attributes were quantified using TLS. Paired sample T-tests showed that Dg, Hg, and G estimated at T1 significantly (p < 0.000) differed from the respective attributes estimated at T2 (Table 7). In case of TPH the differences between T1 and T2 estimates were not considered significant (p > 0.05). Based on field measurements, TPHT1 was significantly larger than TPHT2 on Norway spruce-dominated and mixed-species sample plots due to fallen or harvested trees. On these sample plots, some of the trees that were fallen or harvested were not detected at T1 and thus, change in TPH was not captured. On Scots pine-dominated sample plots, however, the change in TPH was not considered significant and thus, not expected to be captured with TLS. In general, accuracy of characterizing changes in forest structural attributes was at a higher level on Scots pinedominated sample plots than on Norway spruce-dominated and mixed-species sample plots (Table  6).
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 8 July 2020 doi:10.20944/preprints202007.0154.v1 Table 7. The p-values from the paired-sample T-tests indicating the significance of the differences between the TLS-derived estimates on forest structural attributes, such as basal area-weighted mean diameter (Dg) and -height (Hg), mean basal area (G) and trees per hectare (TPH) by main tree species measured in 2014 (T1) and 2019 (T2).

Discussion
The objective of this study was to assess the feasibility of TLS in characterizing boreal forest structure in space and time. We used a bi-temporal TLS dataset covering a five-year growth period in between the data acquisition campaigns in 2014 (T1) and 2019 (T2) and analysed the accuracy of the TLS-based method to quantify changes in tree and forest structural attributes. The results showed that changes in tree and forest structural attributes were captured using TLS. In general, tree and forest structural attributes estimated at T1 differed significantly (p < 0.01) from the respective estimates at T2 (Tables 5, 7). Only in case of TPH the differences in the TLS-derived estimates at T1 and T2 were not considered statistically significant due to incomplete tree detection. In general, changes in the tree attributes were estimated more accurately for Scots pine trees, followed by Norway spruce and broadleaved trees (Table 4). Similarly, the accuracy of characterizing changes in forest structural attributes was higher on Scots pine-dominated sample plots than on Norway sprucedominated and mixed-species sample plots ( Table 6).
The accuracy of using TLS in estimating tree and forest structural attributes at T1 and T2 was comparable to the previous findings on the performance of TLS in forest characterization. Based on previous studies, it is known that the performance of a TLS-based approach to characterize forest structure is affected by 1) the ability of the method to detect trees, and 2) the accuracy of the method to estimate tree attributes [37,41]. With our method, the g of trees that were detected at T1 and T2 accounted for 84.5% of the basal area of all trees ( Table 3). Trees that remained undetected were mainly Norway spruces and broadleaved trees, and small in dbh ( Figure 2). Similar performance in tree detection was reported in comparable forest conditions by Yrttimaa et al. [41]. Successful tree detection and characterization requires that a tree must be sufficiently represented by a point cloud. Occlusion decreases the capacity of a point cloud to properly digitize trees or forest stands [42]. Use of a multi-scan approach in TLS data acquisition decreases point cloud occlusion, thus improving the performance of a TLS-based method to characterize forest stands, especially the ones with a complex structure [43][44][45]. On a forest stand, occlusion affects the performance of a TLS-based method on both horizontal and vertical dimensions. Horizontal occlusion causes undetection of trees, as a part of trees remain shaded by other trees and undergrowth vegetation [41,42]. Differences in the tree detection accuracies between the different tree species can be explained by the different shapes of dbh distributions. Small trees have less surface visible to the scanner than large trees and thus, are more likely to be completely occluded by other trees (Figure 2). Undetection of trees directly affects the accuracy of G and TPH estimates which are, by definition, computed by summing up individual tree attributes at the sample plot level (see Table 2). In this study, G and TPH were underestimated by 19.1-20.5% and 26.9-35.0%, respectively, with RMSEs of 26.9% and 42.9-50.5%. These results are comparable to the accuracy achieved in similar forest conditions by [41]. In general, attributes characterizing horizontal tree dimensions of forest structure were estimated with high accuracy. In this study, dbh of all trees was estimated with an RMSE of 0.90-1.18 cm (4.1-5.7%) which is similar to results obtained by earlier studies in boreal forest conditions [20,43]. For Dg the estimation error (RMSE 1.42-1.60 cm / 5.2-6.0%) was slightly smaller than in [41] where an RMSE of 3.1 cm (12.3%) was recorded in comparable forest conditions. Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 8 July 2020 doi:10.20944/preprints202007.0154.v1 Vertical point cloud occlusion causes uncertainty in h estimates as dense canopy layers and overlapping crowns block the visibility from a scanner to treetops [8,46]. Thus, performance of TLSbased approaches to correctly estimate h has been identified as a major bottleneck hindering the use of TLS in practical forest inventory applications [47]. This was confirmed also in this study, as h was underestimated due to limited visibility to treetops. An RMSE of 4. 10-4.37 m (19.7-22.5%) in h estimates was recorded for all the trees which is in line with previous findings in comparable forest conditions [43,47]. As expected based on results reported in [41], accuracy of characterizing forest height was improved at the sample plot level where an RMSE of 1.85-2.51 m (7.9-11.2%) was recorded for Hg estimates. For Scots pine trees, hc was successfully estimated by investigating the outer dimensions of a horizontally binned point cloud consisting of non-stem points (see Figure 1). Accuracy in hc estimates was similar to h estimates of Scots pine trees, providing reliable estimates for cr ( Table 4).
The absolute errors in quantifying changes in tree attributes were, in general, at the same level than the errors in characterizing the attributes at T1 and T2 (Table 4). However, the bi-temporal TLS data used in this study covered only five growing seasons, which is a very short time when considering the lifetime of Scots pine or Norway spruce trees normally reaching the age from 80 to 120 years in boreal forests. Thus, changes in tree and forest structural attributes were relatively small, which explains why the errors in quantifying changes in tree attributes were rather large relative to the field-measured changes. In this study, for example, RMSE in dbh estimates (0.90-1.18 cm) was similar to the recorded average increase in dbh (1.16 cm). Short monitoring periods have been still applied in the previous studies related to investigating changes in forest structure using TLS due to the novelty of the method. In this case, it needs to be noted that the effect of measurement errors play an important role when the changes in the monitored attributes were small. This means that, for example, the recorded changes in dbh and h can be within the expected errors of both TLS-based and traditional measurement methods reported by [8,48]. As Luoma et al. [48] showed earlier, there is some variation also in the calliper and clinometer measurements when the measurements are repeated, which can depend on the different measurement positions (e.g. determining the breast height for dbh measurements) among other subjective factors. Dendrometers, such as a girth band for tree diameter measurement [49], provide probably the most accurate observations for monitoring changes in the dimensions of a living tree. However, dendrometers only measure one attribute at a time (e.g. dbh) and installing girth bands on a large number of trees to only monitor changes in dbh at forest stand level is rather expensive. Thus, data from clinometer and calliper measurements is typically used as a reference for tree attributes. But with their reliability being on a similar level to TLS-derived estimates, especially when measuring tree stem diameters, it needs to be remembered that actually either one of the measurements could be the ground truth, or the true value may be somewhere between the two observations. With longer monitoring periods, changes in tree attributes are expected to be larger and thus, the effect of the measurement errors should decrease, which further improve the reliability of change monitoring. Despite the variation in accuracies, the results of this study allow us to expect that with longer time periods, also the relative accuracies in change monitoring will be improved when the changes in tree attributes to be captured also increases..
There are only a few previous studies where TLS has been used to capture changes in tree attributes. Liang et al. [6] reported that TLS-derived bi-temporal changes accounted for 92 percent of the changed basal area due to harvesting. Sheppard et al. [32] detected changes in the structure and form of wild cherry trees. Kaasalainen et al. [31] and Srinivasan et al. [30] investigated changes in biomass which were either studies in laboratory condition [31] or modelled [30], but not in forest conditions or directly. All the earlier studies only concentrated on a few attributes with a relatively small number of trees, hindering comparisons between this and previous studies. Luoma et al. [35] tested the feasibility of detecting tree growth with TLS on 35 sample trees with a time period of 9 years and made similar conclusions on the abilities of TLS in detecting changes in tree attributes. Whereas all the measurements in this study were TLS-based, Luoma et al. [35] still used fieldmeasured h in their study. Based on our knowledge, this study is the first attempt to investigate changes in tree and forest structural attributes using TLS, which has been performed with such a large number of trees and in varying forest conditions. Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 8 July 2020 doi:10.20944/preprints202007.0154.v1 The results of this study confirm the feasibility of TLS in characterizing forest structure in space and time. If an increase or decrease in tree and forest attributes was recorded in the field with conventional mensuration tools, a similar outcome was achieved by using bi-temporal TLS data. TPH was the only attribute that could not be characterized in time due to incomplete tree detection at both timepoints. Changes in TPH were minor, mostly because of a couple of fallen or harvested trees per sample plot during the monitoring period. Thus, it is evident that all trees should be characterized at the beginning and at the end of the monitoring period to reliably estimate ΔTPH. This requires better performance in tree detection which could be achieved with a more complete coverage of TLS data to avoid horizontal point cloud occlusion. However, increasing the number of individual scans used in a multi-scan approach decreases the cost-efficiency of TLS data acquisition. Use of mobile laser scanning (MLS) instead, where a point cloud is collected with a laser scanner mounted on an allterrain vehicle [50,51] or a backpack [51,52], or with a hand-held laser scanner [53,54], would presumably be a more cost-efficient option to cover entire forest stands. When combined with SLAM (Simultaneous Localization and Mapping) -technology [55], the accuracy of MLS-derived tree attribute estimates is expected to be close to the TLS-derived estimates [52,54]. Another mobile platform for a laser scanner to suit close-range forest monitoring is an unmanned aerial vehicle (UAV) which can be used to collect detailed point clouds from above a forest canopy [56][57][58]. Due to different data acquisition geometries between terrestrial and UAV-borne point clouds, UAV-borne laser scanning is more suitable for characterization of the vertical forest structure whereas TLS or MLS can better capture the horizontal forest structure. An alternative option to combine the benefits of both terrestrial and aerial point cloud-based approaches is to collect the UAV-borne point clouds from inside the canopy [59] or to use a multisensorial approach [60] where photogrammetric UAV point clouds were integrated with TLS data. Use of a combination of bi-temporal terrestrial and aerial point clouds is expected to improve the accuracy of vertical forest characterization in space and time.

Conclusions
It is known that TLS is capable of characterizing forest structure in detail in space. Thus far, there has been a limited understanding on how forest structural changes can be quantified using TLS in time in boreal forest conditions. The results of this study confirm the capacity of TLS in providing information on the changes in tree and forest structural attributes. If an increase or decrease in tree and forest attributes was recorded in the field with callipers and clinometers, a similar outcome was achieved by using bi-temporal TLS data. However, incomplete digitization of trees and forest stands due to vertical and horizontal occlusion causes uncertainty in TLS-derived estimates in tree and forest attributes and their changes. Vertical occlusion could be decreased by using a combination of terrestrial and aerial point cloud data. Horizontal occlusion could be decreased with more complete point clouds by applying MLS technique to preserve cost-efficiency in data acquisition.
In this study, changes in tree and forest structural attributes were small due to a relatively short monitoring period. It is expected that, with a longer monitoring time the changes in tree attributes become more reliably detectable when automated point cloud-based approaches are used in boreal forest conditions.