Detecting shrub encroachment in seminatural grasslands using UAS LiDAR

Abstract Shrub encroachment in seminatural grasslands threatens local biodiversity unless management is applied to reduce shrub density. Dense vegetation of Cytisus scoparius homogenizes the landscape negatively affecting local plant diversity. Detecting structural change (e.g., biomass) is essential for assessing negative impacts of encroachment. Hence, exploring new monitoring tools to achieve this task is important for effectively capturing change and evaluating management activities. This study combines traditional field‐based measurements with novel Light Detection and Ranging (LiDAR) observations from an Unmanned Aircraft System (UAS). We investigate the accuracy of mapping C. scoparius in three dimensions (3D) and of structural change metrics (i.e., biomass) derived from ultrahigh‐density point cloud data (>1,000 pts/m2). Presence–absence of 12 shrub or tree genera was recorded across a 6.7 ha seminatural grassland area in Denmark. Furthermore, 10 individuals of C. scoparius were harvested for biomass measurements. With a UAS LiDAR system, we collected ultrahigh‐density spatial data across the area in October 2017 (leaf‐on) and April 2018 (leaf‐off). We utilized a 3D point‐based classification to distinguish shrub genera based on their structural appearance (i.e., density, light penetration, and surface roughness). From the identified C. scoparius individuals, we related different volume metrics (mean, max, and range) to measured biomass and quantified spatial variation in biomass change from 2017 to 2018. We obtained overall classification accuracies above 86% from point clouds of both seasons. Maximum volume explained 77.4% of the variation in biomass. The spatial patterns revealed landscape‐scale variation in biomass change between autumn 2017 and spring 2018, with a notable decrease in some areas. Further studies are needed to disentangle the causes of the observed decrease, for example, recent winter grazing and/or frost events. Synthesis and applications: We present a workflow for processing ultrahigh‐density spatial data obtained from a UAS LiDAR system to detect change in C. scoparius. We demonstrate that UAS LiDAR is a promising tool to map and monitor grassland shrub dynamics at the landscape scale with the accuracy needed for effective nature management. It is a new tool for standardized and nonbiased evaluation of management activities initiated to prevent shrub encroachment.

efforts that highlight the need for action (e.g., shrub reduction) should be strategically aligned with the actual spatial scale of management (Magurran, 2016). However, monitoring shrub dynamics with the detail (i.e., small grain size) and spatial extent needed for management is challenging (Cao, Liu, Cui, Chen, & Chen, 2018).
Remote sensing covers larger areas than classical field assessments of structural change, which are time-consuming and provide only local information (Wachendorf, Fricke, & Möckel, 2017).
Satellite-based spectral information can provide valuable information on cover of C. scoparius at the landscape scale for areas with high shrub densities, although limited to observations in the flowering period (Hill, Prasad, & Leckie, 2016). However, the relatively low resolution of satellite imagery (Aplin, 2005) does not allow the quantification of structural information, for example, height and biomass, from individuals or lower concentrated areas of shrubs.
LiDAR technology is a remote sensing method providing three-dimensional (3D) point cloud information suitable for quantifying vegetation structure (Lefsky, a, Cohen, W. B., Parker, G. G., & Harding, D. J., 2002). Airborne LiDAR has been used to map grassland vegetation (Zlinszky et al., 2014) and to explain variation in species diversity across varying vegetation communities (Moeslund et al., 2019).
Especially, forestry research has successfully implemented airborne LiDAR data to, for example, detect the composition of gymnosperm species in plantations (Donoghue, Watt, Cox, & Wilson, 2007), to evaluate ecosystem services (Vauhkonen, 2018) and management efforts based on forest structure (Valbuena, Eerikäinen, Packalen, & Maltamo, 2016), or to demonstrate how LiDAR can be used to measure vegetation height of sagebrush (Mitchell et al., 2011).
Furthermore, stages of shrub encroachment and biomass estimates have been mapped on a coarser resolution (30 m raster) based on a LiDAR point density of 5.6 points/m 2 (Sankey, Shrestha, Sankey, Hardegree, & Strand, 2013). With advancing technology, it is now possible to mount good quality scanners on UAS (Manfreda et al., 2018). UAS LiDAR systems provide new opportunities to provide ultrahigh point density (>1,000 pts/m 2 ) on demand; hence, mapping vegetation structure with high detail and sampling frequency becomes possible across areas of up to several square kilometers. UAS LiDAR systems have been used to detect individual trees and to measure metrics such as height and stem diameter (Wallace, Lucieer, Watson, & Turner, 2012;Wieser et al., 2017). The ability to detect single trees has been found to increase with higher point densities (Wallace, Lucieer, & Watson, 2014), and Balsi, Esposito, Fallavollita, & Nardinocchi, 2018 could contrast different shapes of horizontal overlapping trees using information from the whole volume of points in the point cloud. Furthermore, Moeslund et al., 2019 analyzed the potential of using airborne LiDAR-derived metrics, including a biomass measure, to assess the diversity of different organisms (i.e., vascular plants, fungi, lichens, and bryophytes).
We demonstrate that UAS LiDAR is a promising tool to map and monitor grassland shrub dynamics at the landscape scale with the accuracy needed for effective nature management. It is a new tool for standardized and nonbiased evaluation of management activities initiated to prevent shrub encroachment.

K E Y W O R D S
biomass, grassland dynamics, remote sensing, scotch broom, shrub encroachment, UAS LiDAR The aim of this study was to assess the value of UAS LiDAR for monitoring structural change in a seminatural grassland threatened by encroachment of C. scoparius. We developed a semiautomatic workflow to measure structural features of different shrubs and to enable standardized monitoring for estimating spatiotemporal changes of C scoparius biomass. Specifically, we addressed the following questions: (a) How accurately can C. scoparius be classified in UAS LiDAR point clouds? (b) How precise can the biomass of C. scoparius individuals be estimated from point clouds? (c) How does the estimated C. scoparius biomass change between two surveys at different seasons?
The area is situated in a temperate climate zone with a mean annual temperature of 7.5°C and mean annual precipitation of 585 mm (Fick & Hijmans, 2017). The terrain elevation varies between 30 and 60 m above sea level. Graminoids and small, broad-leaved herbs characteristic for European dry grasslands dominate the vegetation, while single standing trees and different species of shrubs are patchily distributed across the area. Cytisus scoparius forms dense stands in parts of the area and a broad-leaf forest is located toward the northeast. Besides naturally occurring wildlife, the area is grazed by Galloway cattle and Exmoor ponies and thus affected by all year grazing and trampling.
The grazers were introduced in 2016 as part of a rewilding experiment for creating a more self-regulating ecosystem (Svenning et al., 2016).

| UAS LiDAR system
We integrated the Surveyor laser-scanning system (YellowScan, Montferrier sur Lez, France) on an 8-rotary wing UAS (MK8-3500; Mikrokopter, HiSystems GmbH; see Figure 1). The Surveyor is a dual-return system ranging in the 903 nm wavelength with 360 degrees scanner angle, based on the Velodyne VLP-16 "Puck" laser scanner with a maximum measurement range of 100 m and a ranging accuracy of 3 cm. The LiDAR sensor payload weighs 1.6 kg. The Surveyor has a global navigation satellite system (GNSS) receiver and inertial measurement unit (IMU) integrated that acts as a rover (Applanix APX15). We utilize a Trimble base-rover postprocessed kinematic (PPK) solution to gain subdecimeter accuracy for the system in XYZ directions (Chaponnière & Allouis, 2016). To validate the estimation of biomass from C. scoparius, ten individuals were harvested for dry weight measurements after 2 days of incubation at 60°C. Before cutting the shrubs, they were measured with the RTK GNSS device, resulting in small manual point clouds of 25 points each (Appendix 3).

| Processing workflow
We developed a workflow for processing ultrahigh-density point clouds from UAS LiDAR to detect and map structural change in shrubs ( Figure 2). Point cloud data processing was performed separately for 2017 and 2018 data with OPALS software v. 2.3.1 specifically developed for handling airborne LiDAR data (Pfeifer, Mandlburger, Otepka, & Karel, 2014).

| Preprocessing and quality control of LiDAR data
The position data recorded during the flight by the LiDAR system were postprocessed using PosPac UAV v. 8.2 with data recorded We performed a relative adjustment of the point cloud to improve the alignment of objects (i.e., vegetation). The procedure was based on fitting overlapping flight strips with a least squares matching technique as described in Ressl, Mandlburger, and Pfeifer (2009). This process aligned the point clouds better relatively by, for example, reducing the variation in ground point height. However, because of the variation in flight strips, the absolute accuracy decreased to a vertical RMSE of 7.6 cm and 9.1 cm for 2017 and 2018, respectively, and a horizontal RMSE in 2018 of 6.5 cm. For further details on preprocessing and quality control, see Appendix 1.

| Applying thresholds to obtain a shrub layer point cloud
We coarsely divided the point cloud into height-based vegetation classes and a ground class to reduce processing time for the later shrub classification. A digital terrain model (DTM) derived from the minimum point height within a 1x1 m moving window was used to calculate normalized height of points (NormZ) (Appendix 2). We then classified the points using thresholds into ground (NormZ < 0.15 m), low vegetation (NormZ < 0.3 m), and high vegetation (NormZ > 3.5 m) leaving a shrub layer point cloud of medium height vegetation (0.3 m < NormZ < 3.5 m).
Calluna vulgaris fell into the low vegetation layer and therefore was excluded from the further shrub classification procedure.

| Generating point-based reference data of structural features
From the 180 measured GNSS points, we gathered adjacent LiDAR points within a size-specific area to comprise a reference dataset of 13 classes (11 shrub taxa, fenceposts, and shrub absence points; Appendix 2). Reference data were collected as presence/absence (PA) of shrubs in late 2017 and early 2018, that is, between the UAS flights. Within this period, we would only expect shrubs to disappear, for example, browsing/trampling by animals, which would mean that such individuals will be included as a GNSS record in 2017 but would not be there in 2018. Because of the digital assignment of training/ validation data, no points would represent the shrub in 2018, and therefore, no points will be assigned to the given shrub class.
We developed structural features derived from the UAS LiDAR point cloud to represent vegetation morphology with the aim to classify shrub species (Appendix 2). We targeted the variables to represent shrub species on varying levels of scale, ranging from small-scale leaf characteristics to larger-scale shrub growth form and shape. The shrub growth form is here interpreted as the general appearance of the shrub and of how branches and leaves are arranged. Hence, it corresponds to specific, although overlapping features, such as structural complexity, density, and light penetration (Popescu, 2011).
Specifically, we calculated the variance in height from a fitted normal-based plane (Pseudowaveform) to resemble roughness or structural complexity (Van Aardt et al., 2012). The amount of points (Point count) and average distance between points (Point distance) were used as a measure of density, while light penetration was represented by including the count of ground points classified by the above mentioned threshold (Ground points count) and the average number of returning echoes (No. of echoes). Additionally, we calculated a Rank feature, where points within the search radius are ranked by lowest to highest point and assigned a corresponding value between 0 and F I G U R E 2 Flowchart of UAS LiDAR processing chain to derive structural information (biomass) from classified shrub species and comparing between two time periods 100. This measure was intended to represent shrub shape and can particularly be useful for recognizing overhanging canopies. Also, characterizing shrub leaf and branch features, the angles between a point and all its neighbors within a search radius were extracted and averaged as the negative openness for each point. The negative openness refers to a conical view looking downwards, while oppositely the positive would be pointing up. It were originally developed for pixel-based terrain modeling by taking the mean openness angle from the eight neighboring pixels (or more, depending on the search radius) in each cardinal and intercardinal direction (Yokoyama, Shirasawa, & Pike, 2002). However, all points in any possible direction within the search radius were considered for the 3D point attributes and therefore not necessarily restricted to eight directions.
From the initial set of 17 structural features, we retained seven after testing for autocorrelation and variable importance (Table 1).
Variable importance was calculated manually based on the decrease in overall accuracy when leaving one variable out in the classification process (Appendix 2). Variables with high pairwise correlation (ρ > .75) and low importance scores were removed from the dataset.

| Machine learning for 3D classification of Cytisus scoparius and other shrub species
For the 3D classification of the shrub point cloud, we utilized the built-in classification procedure in the OPALS software. It uses the tree-based decision algorithm (De'ath & Fabricius, 2000) termed recursive partitioning via the rpart package for R (Therneau & Atkinson, 1997). The algorithm operates by dividing the dataset for each chosen variable separately, that is, it finds the splitting value for tree branches which results in the purest nodes, that is, most homogenous. A perfectly pure node refers to all observations being assigned the correct label caused by the split value. When the variables could not further increase the node purity, the resulting decision tree was pruned with the complexity parameter set to 0.001 for a final simplified decision tree. We fixed the complexity parameter after a trial process where we lowered the parameter stepwise until the accuracy started to decrease. The aim was to run the classification with the lowest possible complexity parameter.
The classification accuracy was assessed by randomly stratifying 90% of the reference data as training and 10% as testing data. The quality and reliability of accuracy assessments are affected by the reference data input and sampling strategy (Millard & Richardson, 2015). We therefore also assessed a 70% training to 30% validation data split (Appendix 2). Stratification was done among the defined classes. We generated 100 such 90/10 and 70/30 training and validation datasets as input for the OPALS classification algorithm.
For each set, classification accuracy was assessed from the resulting confusion matrices using R to calculate overall accuracy and Kappa values, while the class-wise accuracies were evaluated with precision, recall and the harmonic mean between the two, termed "F1" . The 70/30 split resulted in <0.5% decrease in overall accuracy and Kappa coefficient (Appendix 2). To increase the amount of training data used in the final model for extracting biomass metrics, we applied a model validated by the 90/10 split (Appendix 2). From the predictions, we obtained the class probabilities allowing fuzzy classifications, that is, the membership of a class is represented by a probability value between 0 and 1 rather than a Boolean value (true or false) as with traditional hard-boundary classification (Foody, 1996;Zlinszky & Kania, 2016). Finally, a classification was performed with 100% of the reference data and printed into the full point cloud for visualization purposes. Additionally, we applied a similar classification procedure as described above, but with the variables projected into 2D raster data (Appendix 2). The 2D classification was performed with R statistics 3.5.0 (R Core Team, 2016) allowing us to test alternative classification algorithms. ) via the lm() function in R. We used R 2 adj values from the models to cope with the relatively low sample size of 10 and thereby avoid making too optimistic conclusions. We applied the model coefficients ( Figure 3) from the best fit (maximum NormZ) to calculate a biomass estimate for the rasterized variable across a 6.7 ha area and for both datasets (2017 + 2018). During the raster projection, we incorporated class probabilities from the final fuzzy classification models (see accuracy assessment in Appendix 2) to exclude points classified as C. scoparius with <60% probability for one set of maximum NormZ values. Likewise, a second set of values was extracted by adjusting this probability threshold more strictly to 80%. At last, the change in biomass was mapped and aggregated from 5 cm resolution to 2.5 m grid cells to emphasize the change.

| Detection of Cytisus scoparius in a 3D landscape of points
After georeferencing and noise filtering, our UAS LiDAR system gen- Furthermore, it is noticeable that C. scoparius shrubs were detected underneath the forest canopy as well (Figure 4a1).

| Cytisus scoparius biomass estimation
The LiDAR-derived volume metrics of NormZ (average; maximum; range) correlated well with the biomass measurements of the 10 harvested C. scoparius shrubs (Spearman's ρ = .87; .88; .88). However, the inclusion of very small shrubs in the harvested samples challenged the LiDAR detection of biomass (Appendix 3). The maximum volume resulted in the best linear fit with R 2 adj = .77 ( Figure 3). The mean and range metrics performed worse in a linear model with R 2 adj = .72 and R 2 adj = .60, respectively. The accuracy assessment from the specific model used for obtaining biomass metrics is presented in Appendix 2.

| Cytisus scoparius biomass change
We extracted the total biomass sum from the points with >60% and >80% probability of being C. scoparius. This resulted in 7,500.4 and 5,257.6 kg in 2017 and 2018, respectively, for >60% and, likewise, 5,320.9 and 4,993.0 kg for >80%, in an area of 6.7 ha. For the comparison, we included only the overlapping areas that had an average point distance below 3 cm. On the landscape scale, this resulted in an average biomass decline of C. scoparius from autumn 2017 to spring 2018 of 33.4 and 4.9 g/m 2 for the 60% and 80% probability thresholds, respectively. However, on a local scale the distribution of biomass changes in the area varied but was similar for both thresholds. An upscaled visualization from 5 cm resolution to grid cells of 2.5 m × 2.5 m revealed a pattern of larger decreases in biomass to be identified in especially the northeastern part of the area, while in other parts, we observed no or a slight increase in biomass between the 2 years ( Figure 5).

| D ISCUSS I ON
With the sole use of UAS LiDAR-derived structural information, we identified shrub classes with an overall mean accuracy of more than 86.9% regardless of the year and detected C. scoparius with at least 96.2% F1 accuracy in the point cloud (Table 2). Furthermore, using a simple volume metric (NormZ) from the classified C. scoparius point cloud, we explained 77% of the variation in actual harvested biomass ( Figure 3) and quantified a reduction of 327.8 kg during the winter period from autumn 2017 to spring 2018, assuming the 80% probability threshold to be most accurate.

| LiDAR-derived structural features for shrub classification
Our study demonstrates that ecologically meaningful features can be extracted from UAS LiDAR point clouds which represent the structural appearance (i.e., growth form, shape, and leaf/branch orientation and arrangement) of different shrub species, especially C. scoparius. We targeted the variables to represent shrub species on varying levels of scale, ranging from small-scale leaf characteristics to larger-scale shrub growth form, density, and shape. For example, C. scoparius is characterized as being light competitive (Peterson & Prasad, 1998), hence, reducing the available light beneath the canopy. However, compared with, for example, the compact canopy of Juniperus communis, it may allow more light to travel through, especially during leaf-off periods, eventually resulting in more ground hits of the LiDAR beams.
The variable that explained most of the variation, Pseudowaveform, was computed with a relatively large search radius (0.5 m) and therefore represents the variation in the structure of shrub branches and leaves (Table 1). By calculating on this relatively larger scale, we TA B L E 2 Accuracy assessment of the point cloud classifications of Cytisus scoparius from autumn 2017 and spring 2018. Classification accuracies are averages from 100 model iterations of randomly selected training/validation data (90/10% split) and with standard deviations. Overall accuracy and Kappa coefficient evaluate the classification of all classes, while the F1 score  assesses the performance in predicting the C. scoparius class. The last row presents the results from a merged classification including both the 2017 and 2018 point clouds expected the variable to be less affected by differences in leaf abundance between the two flights. From analyzing the shrub signatures, we find Pseudowaveform to be stable between the two seasonal states and throughout the classes, except for Betula (Appendix 3).
Structural complexity in vegetation is often assessed by the variance of derived LiDAR metrics relative to a horizontal plane (Kane et al., 2010;Kukunda et al., 2019). Textural metrics in image analysis can similarly be used to distinguish structurally dissimilar species (Oldeland, Naftal, & Strohbach, 2017), but are limited to detect only 2D surface differences. Here, we utilize information of 3D vegetation structure. The variance for example will be set relative to a  ing a relatively round and smooth canopy, whereas the branch gaps will be more exposed during leaf-off period. Finally, to differentiate between C. scoparius and low tree species in the shrub point cloud with overhanging canopies, for example, Malus sp. or Prunus cerasifera, the classification benefitted from the Rank variable. Here, lower points from a hanging branch without points underneath would be ranked lower than C. scoparius branches, which often will have lower points beneath, from low vegetation or ground. Thus, computing ecologically meaningful variables from 3D point clouds is possible but needs to be based on ecological knowledge and targeted on the focal species or vegetation class.
The performance of the LiDAR-derived structural variables to detect and distinguish shrub species in this study demonstrates a promising use at the level of detail obtainable with a drone-based platform ( Table 2). In addition to the increased spatial resolution and structural detail, our landscape study of C. scoparius differs from Hill et al. (2016) in being independent from observations during flowering periods. However, phenological events can also be considered an important aspect of understanding the development of shrub encroachment. In particular, an increase in temporal resolution would be beneficial for remote sensing studies of vegetation to understand how such variables vary with seasonality. A study by Müllerová et al. (2017) demonstrates an example of this, by recognizing two invasive species from differently scaled images and throughout the season.
Hence, using spectral information in combination with LiDAR could extend the monitoring possibilities even further.

| Workflow and classification challenges
The use of LiDAR-derived features poses many challenges and demands a novel way of developing and understanding these 3D measures in an ecological context. Thus, further improvements are possible by implementing ecological or biological knowledge in the computed variables. Depending on the focal species, the LiDAR-derived features can be adjusted to fit specific morphological characteristics, and a more general approach could be developed to assure transferability for many shrub species. For C. scoparius, the change in structural signature between leave-on and leave-off is expected to be low due to the small leaves (Appendix 2). However, with the onset of flowering, this might change for some of the variables. Alternatively, for other shrub species it might be important to develop variables that are independent of seasonal states to assure transferability of the classifier.
For C. scoparius, we found a significant change in the sample distributions in three of the density variables (PCount_ground, PCount, and PDist) which indicates that seasonal variability might be seen in these variables. However, it is the combined signature that is important for the classification process and the distinction among species.
When evaluating normalized vegetation height (NormZ) used for biomass estimation, we find no difference between the two flight dates in C. scoparius, which is different for Betula species. These findings highlight the need for variable adjustments and/or seasonal timing, to obtain reliable biomass estimations for a given target species.
We implemented a workflow to improve the assessment of shrub biomass during encroachment. During the development, we experienced several challenges which could affect the results and need careful consideration and evaluation during implementation: (a) misalignment reduce spatial accuracy, (b) thresholding of major Threshold approaches to filter data noise as well as for doing a first general classification are essential for effective data processing.
However, finding meaningful thresholds in heterogeneous plant communities with highly variable structures is difficult and hard to generalize across a larger area. Therefore, deciding on exact thresholding values often is a trade-off issue, by either including an amount of noise or excluding useful data. In our study, we emphasized to include as much data as possible, accepting some levels of uncertainty. Still, with broadly defined threshold values, points representing a target species may disappear in these first steps of filtering and classifying. For example, we attempted to include training data from the small-sized shrub Calluna vulgaris; however, these points were set aside in the general class "Low vegetation" and thus not included in the tree-based classification. Reducing the dataset is crucial for optimal processing speed and point cloud visualization but choosing general classes and thresholds must be evaluated and depends on the target species. If targeted, the C. vulgaris shrubs could most likely be identified by running the classification procedure on the low vegetation points and could perhaps be implemented in a future study including herbaceous vegetation.
Generally, one should aim for balanced training/validation data (Millard & Richardson, 2015

| Spatial patterns of C. scoparius biomass change
Depending on the probability threshold for detecting a C. scoparius shrub, we estimated a decrease in biomass change of 4.9-33.4 g/m 2 between the two mapping dates. Nevertheless, we found a consistent spatial pattern of biomass change for both thresholds, indicating a large decline in the northeastern part of the area and no or a slight increase in the southwestern part ( Figure 5). There might be several reasons for this observed spatial pattern: First, the slight increase in biomass could result from late autumn or early spring growth. The shrubs are not expected to grow during a Danish winter season; however, the timing of flight campaigns may still include some late autumn or early spring growth. In addition, C. scoparius branches contain chlorophyll, which might be able to induce growth in warmer periods and favor early growth in spring. The reason why biomass increase was mainly observed toward the southwest of the study area could be low inter-and intraspecific competition or variation in microclimatic conditions. A freestanding individual without neighbors is more likely to grow and expand due to lack of competition when the conditions are suitable. In contrast, competing neighbors may supress growth of C. scoparius individuals in densely populated surroundings. The possible intraspecific interaction has previously been raised in the literature (Paynter, Fowler, Memmott, & Sheppard, 1998), and in accordance with this, the observed spatial pattern in biomass changes suggests that C. scoparius in the open land (midwestern F I G U R E 5 Orthophoto from the study area (Kortforsyningen Danmark, 2019) with overlay of change in maximum Cytisus scoparius volume between 2017 and 2018 UAS LiDAR survey. A C. scoparius class probability threshold is applied and only allowing points classified with >60% (top) and >80% (bottom) probability to be included for change detection. The change is upscaled from 5 cm to 2.5 m 2 for visualization. Blue squares represent growth, while red squares show a biomass decrease. White squares indicate no or little change part) tended to slightly increase, while the denser stands toward the northeast of the study area showed a decline in biomass ( Figure 5).
Second, the reasons for a biomass decline could be either leaf fall (Peterson & Prasad, 1998), which due to the summed 2.5 m grids, is expected to be larger in denser stands, or external factors, such as grazing or frost. Galloway cattle and Exmoor ponies graze the area as part of a rewilding initiative, and no additional food sources are supplied to the animals. Hence, at wintertime when the green vegetation becomes sparse, the animals might feed on hardy shrubs like C. scoparius, or cause damage to shrubs by trampling. Therefore, one alternative hypothesis explaining the spatial pattern of C. scoparius biomass change is that the animals are favoring the northeastern part of the study area.
This area is substantially more forested, and the grazers might benefit from shelter provided by the trees during harsh winter conditions.
Alternatively, abiotic conditions linked to topographic features, such as light availability and freezing temperatures during winter might limit plant growth and harsh winters would potentially cause C. scoparius to die (Peterson & Prasad, 1998).

| Relevance for nature management
The spatial variation in C. scoparius biomass change indicates that shrub dynamics differ in the subareas of the mapped area, possibly due to varying importance of ecological drivers, also yet to be studied ( Figure 5). Our findings highlight that detection of change in shrub density and biomass with high resolution is important when assessing shrub encroachment in monitoring programs for nature management. Traditionally, remote sensing studies have classified vegetation with a two-dimensional approach from spectral information or rasterized LiDAR information. These raster-based methods are efficient in identifying cover or presence of certain grassland species and vegetation types (Hellesen & Matikainen, 2013;Zlinszky et al., 2014). However, to quantify biomass changes and to fully understand the effect of shrub encroachment on plant diversity more comprehensive knowledge is needed. While LiDAR observations from a manned aircraft were applied to map coarse-scaled shrub encroachment characteristics from a single species (Sankey et al., 2013), we are now able to separate species or genera directly in the point cloud and to detect fine-scale biomass dynamics from a target species by utilizing a drone platform. With the use of the established workflow based on a UAS LiDAR system, we provide a new approach for monitoring shrub species dynamics. In this study, C. scoparius is covering the spatial extent corresponding to management operations, but also capturing the local-scale information needed for detecting change and its spatial variation. Again, the benefits of using LiDAR are the vegetation penetration ability (Lefsky et al., 2002) and our findings suggest that LiDAR-derived point clouds are of such a quality, that detecting species of interest even beneath a covering canopy is achievable, as, for example, forest understory species. This will make it possible to monitor and guide management programs for noxious invaders such as Rhododendron ponticum (Sanders, 2017) or rare species such as Allium tricoccum beneficial for indicating favored nature or ecological conditions (Leduc & Knudby, 2018 (Kesting et al., 2015).

| CON CLUS ION
Our study presents a novel method for assessing shrub dynamics in 3D, based on the arrangement and orientation of points in space.
We demonstrate an efficient way of determining specific structural

ACK N OWLED G M ENTS
BM and the remote sensing equipment were funded by a Carlsberg

CO N FLI C T S O F I NTE R E S T
The authors declare no conflict of interest.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data used in this study are available through the Dryad Data  strip data to be further processed. These strips were afterwards readjusted, to make sure that data gathered while the UAS was turning were excluded. At last, data were delimited to a scan angle of ±55 degrees minimizing the amount of noisy points in the dataset.

Quality control
Quality was assessed by creating point density maps to evaluate the coverage throughout the area of interest ( Figure A1.1a,b). Likewise, (New York, N.Y.), 303 (5665)  the dislocation differences between the overlapping flight strips were calculated and visualized on raster maps ( Figure A1.1c-f) as well as in cross-sections of the point cloud ( Figure A1.2). Thickness of the ground points layer was evaluated by measuring a cross section in a place with a minimum of vegetation (i.e., ground thickness should be low) using the point measurement tool in CloudCompare (v2.10, GPL software, 2019, retrieved from http://www.cloud compa re.org/).

Relative adjustment of point cloud
It is possible to reduce the differences between overlapping flight strips (i.e., relative accuracy) caused by system inaccuracies (e.g., These errors are visible as rectangular artifacts in Figure A1.1c-f

F I G U R E
and as differently colored layers in Figure A1.

Outliers
Outliers in the point cloud, that is, erroneous records of LiDAR returns, will affect further processing and the quality of point classification.
Hence, such outliers have been deleted by attributing each point with the number of points within a search radius of 1 m, that is, a sphere with a diameter of 2 m around the given point and then applying a two-step procedure. First, we deleted all points that had no neighbors within the 1 m search radius, which corresponds to single noise points.
Second, as noise can appear in clumps we also deleted all points with average number of neighbors <3 from points within a search radius of 3 m, that is, a sphere with a diameter of 6 m around a given point.

Absolute precision
We measured 12 fencepost-tops with a differential GNSS system yello wscan -lidar.com/produ cts/surve yor/). However, the accuracy of the absolute position of these control points decreased with deviations of 6.5 cm and 8.3 cm, respectively, after the relative geo-referencing of flight strips, that is, relative adjustment of the point clouds ( Figure A1.3).
In addition, we measured the systems XY precision in the 2018 survey by using five ground control markers. Here, we found the deviation of 1.1 cm to increase to 4.7 cm after the relative adjustment ( Figure A1.4).
Using these measurements, we calculated the root mean square error (RMSE) and mean absolute difference (MAD), which are presented in Table A1.1.
At last, we overlaid the point clouds from each year to visualize the consistency of the two surveys ( Figure A1.5). It is noticeable that the red points (2017) seem a bit higher located for ground points most probably a vegetation effect in the leaf-on period.
Overall, the red and blue points are still in seemingly good aligned/ intermixed.

Concluding remarks
Our findings indicate that the UAS LiDAR system used for this study performed with an absolute position error below 10 cm in both, vertical and horizontal directions. The absolute accuracy is lower than reported in the technical specification for the LiDAR system; however, the assessment is based on the position of fencepost which has been measured with a GNSS RTK system providing ≤2 cm precision.
Furthermore, we applied a flight strip adjustment procedure that increases the relative accuracy of the point clouds on the expense of slightly shifting the point clouds in absolute geographic space.
Precise alignment of the point clouds was, however, considered most important for the overall purpose of classifying shrubs.  (Waldhauser et al., 2014). The classification workflow targets the recognition of Cytisus scoparius shrubs in a highly heterogenous and vegetated area of semi-natural grassland.

R E FE R E N C E S
The workflow in this appendix is described in a three-step procedure of (a) preparing and training the model, (b) computing and evaluating input attributes, and (c) final classification results.

Preparing and training the model
Prior to the machine learning process, the full point cloud was di-

Training a tree-based model
Two separate sets of reference data were collected in between the two flight surveys. They consist of single differential GNSS measurements of different objects in the area. In the period between reference data and LiDAR flights no major change to, for example, shrubs are expected and following the described method below for generation of validation points, it should be safe to use the two datasets individually and together.
We measured in total 180 shrub individuals from 12 different shrub/tree taxa with an RTK GNSS device, to create the dataset for training and validation. The reference data gathering was performed in random transects throughout the area and spaced approximately 50-100 m. Along each transect every 100 m, the nearest shrub of each species in the area was sampled, where each shrub was meas-

Generating training and validation points
Because Opals modules for attribute assignments are not always working with shapefiles the reference data needed to be pre-processed for implementation in an Opals data manager file (ODM).
Buffers of 25, 35, and 50 cm were created around the reference datapoints in QGIS and selected appropriately to cover a single object of each class. The 50 cm buffer was used for all shrub classes except Calluna vulgaris, which would be decreased to 35 cm buffer radius along with the Not Shrub class. The fenceposts themselves are varying in radius between 10 and 20 cm, and as they would not always be completely straight, the 25 cm buffer was used here.
In QGIS, the buffered reference points were rasterized to 5 × 5 m GeoTIFF's for each class displaying the reference cells with 1 and the rest of the flight area with 0.
Having the reference information in raster files allows it to be stored as attributes in the working ODM with OpalsAddinfo.
Afterwards, all points overlapping in 2D with the raster cells equal to 1 could be assigned as validation points in the given class. To ensure no classes overlapped in the 2D buffer area, the ground, low vegetation, and high vegetation classes were assigned separately.
This method of generating validation points in a point cloud ensures that no false points are introduced in the training model as the actual points are measured with the LiDAR system itself, and the reference data used as a guide for assigning the points.

Input attributes
The tree-based classification procedure available in the OPALS software package is applied on a point-based level, meaning that each point is classified independently from other points. However, before the classification process is initiated, each input attribute was averaged within a cylinder with radius 0.25 m and height ±2.5 m from every point (see Figure A2.2). In this way, the "salt and pepper effect," known also from image classification (Blaschke, Lang, Lorup, Strobl, & Zeil, 2000), is avoided and the scale of the input attributes is more likely to correspond to the size of the target shrubs.
During the averaging process, the point-cloud was filtered for also referred to in this document . For evaluation, the accuracy measures were summed without weight and ranked from best to worst ( Our findings regarding shrub signatures reveal many new possibilities in terms of understanding vegetation dynamics but would require a much more focused effort for data collection which was beyond the scope of the present study.

Final classification results
We

Classification accuracy assessment
We assessed the accuracy of the final classification input by performing a randomly stratified selection of input training/valida-

S H R U B C L A S S I F I C A T I O N
Bell, 1997). The Kappa coefficient compares the observed accuracy with the expected accuracy by random chance. For class-wise accuracy, we used precision (P) and recall (R) values to calculate the harmonic mean termed "F1" . The F1 is calculated by the ratio of (2 * recall * precision) and the sum of recall and precision.
For the definition of OA, recall (i.e., completeness) and precision (i.e., correctness) see above. The results from the accuracy assessment are presented in Figure A2.

Additional 2D classification
We performed a traditional 2D classification based on the rasterized variables for comparison with the 3D process. The entire classification was performed with the "caret" package in R sta- Therefore, rethinking the creation, assignment, and increasing the amount of training data would be necessary to solve this question from a 2D perspective.
The classification results presented in Table A2 F I G U R E A 2 . 6 Density plots showing the resulting class-wise accuracy distribution from 100 permutations of the classification models for the 2017 (Mols17) and 2018 flights (Mols18). The class-wise accuracies are evaluated on the 90/10 split with the F1 value, based on the precision (P) and recall (R) values . Below the density plots, corresponding tables highlight the calculated mean and standard deviation (SD) for each distribution

Data collection
Before cutting the shrubs, the structure and shape were measured with an RTK GNSS device, resulting in small manually created point clouds of each 25 points (Cyt3D in Figure A3.1). The Cyt3D point clouds were constructed by always measuring top and bottom points first, defined as the most upper splitting of branches and the lowest part measurable on the main stem. The additional 23 points were measured in various heights and distances from the main stem, but always on a branch split or directly on the main stem. This was expected to minimize the inclusion of the smaller, and in wind moving branches, which is possibly not detectable by the UAS LiDAR system. In Figure A3.1, the alignment of the manual constructed point cloud to the LiDAR point cloud is visualized to highlight the precision of the different data sources.

Generating and collecting digital biomass data
As described in the classification document (Appendix 2), we clas-