Methodological considerations of terrestrial laser ...

Environ Monit Assess (2017) 189:578 https://doi.org/10.1007/s10661-017-6300-0

Methodological considerations of terrestrial laser scanning for vegetation monitoring in the sagebrush steppe Kyle E. Anderson & Nancy F. Glenn & Lucas P. Spaete & Douglas J. Shinneman & David S. Pilliod & Robert S. Arkle & Susan K. McIlroy & DeWayne R. Derryberry

Received: 8 May 2017 / Accepted: 12 October 2017 # Springer International Publishing AG 2017

Abstract Terrestrial laser scanning (TLS) provides fast collection of high-definition structural information, making it a valuable field instrument to many monitoring applications. A weakness of TLS collections, K. E. Anderson Department of Geosciences, Idaho State University, Pocatello, ID 83209, USA e-mail: [email protected] N. F. Glenn (*) : L. P. Spaete Boise Center Aerospace Laboratory, Department of Geosciences, Boise State University, 1910 University Drive, Boise, ID 83725, USA e-mail: [email protected] L. P. Spaete e-mail: [email protected] D. J. Shinneman : D. S. Pilliod : R. S. Arkle : S. K. McIlroy U.S. Geological Survey, Forest and Rangeland Ecosystem Science Center, 970 Lusk Street, Boise, ID 83706, USA D. J. Shinneman e-mail: [email protected] D. S. Pilliod e-mail: [email protected] R. S. Arkle e-mail: [email protected] S. K. McIlroy e-mail: [email protected] D. R. Derryberry Department of Math, Idaho State University, 921 S. 8th Ave., Stop 8085, Pocatello, ID 83209-8085, USA e-mail: [email protected]

especially in vegetation, is the occurrence of unsampled regions in point clouds where the sensor’s line-of-sight is blocked by intervening material. This problem, referred to as occlusion, may be mitigated by scanning target areas from several positions, increasing the chance that any given area will fall within the scanner’s line-of-sight from at least one position. Because TLS collections are often employed in remote regions where the scope of sampling is limited by logistical factors such as time and battery power, it is important to design field protocols which maximize efficiency and support increased quantity and quality of the data collected. This study informs researchers and practitioners seeking to optimize TLS sampling methods for vegetation monitoring in dryland ecosystems through three analyses. First, we quantify the 2D extent of occluded regions based on the range from single scan positions. Second, we measure the efficacy of additional scan positions on the reduction of 2D occluded regions (area) using progressive configurations of scan positions in 1 ha plots. Third, we test the reproducibility of 3D sampling yielded by a 5-scan/ha sampling methodology using redundant sets of scans. Analyses were performed using measurements at analysis scales of 5 to 50 cm across the 1-ha plots, and we considered plots in grass and shrubdominated communities separately. In grass-dominated plots, a center-scan configuration and 5 cm pixel size sampled at least 90% of the area up to 18 m away from the scanner. In shrub-dominated plots, sampling at least 90% of the area was only achieved within a distance of 12 m. We found that 3 and 5 scans/ha are needed to sample at least ~ 70% of the total area (1 ha) in the grass

578

Page 2 of 12

and shrub-dominated plots, respectively, using 5 cm pixels to measure sampling presence-absence. The reproducibility of 3D sampling provided by a 5 position scan layout across 1-ha plots was 50% (shrub) and 70% (grass) using a 5-cm voxel size, whereas at the 50-cm voxel scale, reproducibility of sampling was nearly 100% for all plot types. Future studies applying TLS in similar dryland environments for vegetation monitoring may use our results as a guide to efficiently achieve sampling coverage and reproducibility in datasets. Keywords Point cloud . Survey . Lidar . Dryland . Vegetation monitoring . Terrestrial laser scanning

Introduction Laser scanning, commonly called light detection and ranging (or lidar), refers to technologies using laser ranging measurements to record precise locational information about targets. Three-dimensional (3D) point clouds yielded by laser scanners are valuable to natural sciences, particularly for measurements and monitoring of vegetation and topography. While many lidar datasets used in the natural sciences are collected from airborne platforms (airplanes, helicopters, unmanned aerial systems (UAS)) (e.g., Andersen et al. 2005; Wallace et al. 2012), recent improvements in technology have led to expanded use of terrestrial laser scanning (TLS) instruments. A TLS instrument consists of a lidar system mounted statically (on a tripod or other stable platform) or dynamically (on a truck or other moving platform, referred to as Bmobile TLS^). Lidar sampling from a TLS offers several advantages, including ultra-high data resolution, below-canopy sampling, horizontal field-ofview and low cost of collection (Heritage and Large 2009). Examples of the rapidly growing body of work using ground-based TLS in vegetation studies includes modeling fuelbeds (Rowell et al. 2016), estimating biomass (Cooper et al. 2017), and inventory monitoring of forests (Strahler et al. 2008), shrublands (Olsoy et al. 2014), and other low stature vegetation (< 2 m) (Loudermilk et al. 2009). A notable disadvantage of scanning from the ground is the issue of occlusion (i.e., Bshadowing^) in point clouds, where the scanner’s line-of-sight to a region is blocked by intervening objects. In aerial collections, a narrow scan angle and field-of-view cause the effects of occlusion to be highly regular,

Environ Monit Assess (2017) 189:578

such that sampling is limited to topmost surfaces and the ability of the laser to penetrate these surfaces. However, the 2D spatial coverage (x, y direction) is largely unaffected by these limitations. In contrast, the mainly lateral scanning orientation of TLS collections may result in large data gaps from vegetation, buildings, or topographic features which occlude regions beyond these features. These data gaps often cause TLS point clouds to be highly irregular, with regions of good sampling coverage falling immediately adjacent to regions that are partially or entirely unsampled. These irregularities of sampling partly explain why the ability to accurately measure vegetation characteristics in previous studies decreases as the distance of the plant to the scanner increases (Greaves et al. 2015). These irregularities can also impede TLS data analysis that uses clustering or interpolation approaches and/or require subsampling prior to analysis (Wilkes et al. 2017, Anderson et al. 2018). Some aerial and terrestrial instruments enable partial mitigation of occlusion through detection of multiple returns of light from each pulse, allowing for sampling of several in-line targets when the first target intercepts only a portion of a pulse’s footprint beam. However, these secondary measurements are at increased risk of pulse paths that reflect off several surfaces and other complications which reduce positional certainty (Shan and Toth 2008). Efforts to mitigate occlusion in TLS data collections for vegetation monitoring will result in a more complete data set for analysis and improve reproducibility of data collections over time and space. Some studies have accounted for occlusion in TLS datasets by modeling unsampled features based on the attributes of sampled features. For example, Strahler et al. (2008) inferred the number of fully occluded tree stems based on the width and density of stems that are sampled, and Henning and Radtke (2006) modeled plant area index in volumetric regions of canopy based on the ratio of reflection to the pass-through of pulses incident to each region, rather than simply measuring the plant area that is actually sampled. Several techniques have been shown to reduce occlusion in TLS collections. Some improvement in sampling coverage may be effected by elevating the instrument well above the objects in the target region, which causes occlusions to fall beneath protrusions rather than stretching out laterally from them. This has been done


using both topographically elevated scanning position sites (Vierling et al. 2013) and elevated scanning platforms (Loudermilk et al. 2009). The most common practice may be to combine scans collected at several positions, increasing the likelihood that a region will fall in the scanner’s field-of-view from at least one position. This technique may be implemented using either a standardized layout of scanning positions or one that is adaptive to local features and has proven useful in sampling extensive plot areas as well as focusing on individual target plants (Clawges et al. 2007; Seidel et al. 2012). A similar technique is the fusion of aerial and TLS point clouds, providing complementary sampling from both above and below forest canopies (Chasmer et al. 2004; Murgoitio et al. 2014). Although many studies employ methods to minimize occlusion in TLS collections, published research that quantifies the efficacy of these techniques is scarce. A small body of literature evaluates the errors in TLSderived elevation models caused by low vegetation occluding the ground surface (e.g., Coveney and Fotheringham 2011; Fan et al. 2014). Van der Zande et al. (2008) studied the extent of unsampled regions as a consequence of occlusion by evaluating the efficacy of two scanning position arrangements (Bdiamond^ and Bcorners^) against a single scanning position at reducing occlusion in simulated forest plots. They found that both arrangements of positions outperformed sampling from a single position, even when the single scan was of a much higher sampling density, but that neither arrangement yielded consistently superior results. Trochta et al. (2013) also tested the number of scans against tree detection in forested plots roughly 1.2 ha in size. They spaced their scan positions 20–110 m apart. They found that by increasing the number of scans, they could improve the detection of trees at longer distances (up to 80% tree detection at 25 m with 3 scans). Cifuentes et al. (2014) found that an orientation similar to scanning from the corners of plots consistently outperformed the Bdiamond^ orientation for estimating forest canopy gap fraction across 20 m × 20 m plots. Cifuentes et al. (2014) also analyzed results from two additional orientations comparable to the Bcorners^ orientation, which yielded inconsistent results and suggested that reproducibility may be related to scan orientation in combination with other variables (e.g., forest stand characteristics, voxel size, and the TLS technology (they used a phase-based scanner)). More recently, Wilkes et al. (2017) present a number of different scan configurations dependent upon

Page 3 of 12 578

forest stem and understory density. Overall, their approach recommends data collection along a chain configuration, where one scan location links to the next. The vegetation structure in dryland ecosystems, including shrublands and grasslands, is particularly well suited for the laterally scanning technology of TLS (for example, recent advances have been made in estimating shrub biomass and leaf area index (LAI) with TLS (Olsoy et al. 2014, 2016)). The efficiency and accuracy of data collections in these ecosystems to estimate vegetation structural parameters will be improved by understanding the limitations of TLS sampling. Knowledge to prevent over- or underexpenditure of sampling effort given study goals, as well as potential mitigation procedures to improve sampling thoroughness, will assist researchers and practitioners. The objectives of this study are to test the extent of TLS occlusion in a dryland (rangeland) ecosystem and to assess the efficacy of methods to mitigate occlusion given the common instrumentation of a laterally scanning instrument mounted on a 2-m base. We designed three experiments to test the following: (1) the effect of scan range on the extent of occluded regions, (2) the effectiveness of several scanning positions at reducing total occluded area in 1 ha plots, and (3) the reproducibility of 3D datasets yielded by a methodology employing 5 scans/ha.

Methods Study area The study area is located in the Morley Nelson Snake River Birds of Prey National Conservation Area (hereafter, NCA, roughly center of 43° 18′ 27.84″ N, 116° 17′ 34.57″ W), which encompasses 240,000 ha of the Snake River Plain ecoregion in southwestern Idaho, USA (Fig. 1a). The NCA climate is characterized by hot, dry summers and mild winters. In an average year, the NCA receives roughly 15–25 cm of precipitation (depending on location), with roughly 38% falling during winter (November– February), 44% during spring (March–June), and 18% during summer to early fall (July–September). Mean July temperature is ~ 24 °C, with maximum summer temperature often exceeding 35 °C, and mean January temperature is roughly − 1 °C (WRCC 2012). Surface geology includes loess

578

Page 4 of 12


Fig. 1 Map of the Morley Nelson Snake River Birds of Prey National Conservation Area (NCA) study area with TLS plot locations (a) and representative examples of shrub (b) and grass (c) plots

windblown soils interspersed by basalt outcrops. The topography consists mainly of plateaus and rolling uplands. The microtopography is generally flat with interspersed pits and mounds caused by burrowing animals. The native vegetation is characterized by a generally open shrub canopy (i.e., < 50% cover), with shrubs up to 1.5 m tall, and an understory of sparse cover of native bunchgrasses (e.g., Poa secunda, Elymus elymoides). Big sagebrush (Artemisia tridentata, primarily ssp. wyomingensis) is the regionally dominant shrub, but other shrub species may be locally dominant and contribute to regional diversity (e.g., Krascheninnikovia lanata). Although wildfire events were historically rare in the NCA, over half of the NCA has burned since 1980. Sagebrush plants are often either slow or unable to recover due to increased fire frequency and competition with fire-adapted, non-native annual species (Brooks et al. 2004; Shinneman et al. 2015). As a result, the contemporary landscape is a mosaic of plant communities, with compositions spanning a gradient between intact native shrublands, shrublands degraded by biological invasion and wildfire, and non-native grasslands where native plants have been replaced by cheatgrass (Bromus tectorum) and other invasive annual grasses and forbs (e.g., Sisymbrium altissimum). Currently, less

than 37% of the NCA retains an intact native shrubland community, in which vegetation cover consists primarily of native grasses and shrubs. Active management on the NCA to promote native flora and reduce wildfire hazard includes strategic grazing, mechanical planting of native species, and mowing (USDI 2008). Data collection Plots (n = 26, each 100 m × 100 m (1 ha)) for field vegetation measurements and TLS data collection were established at locations throughout the NCA, with sites divided evenly between intact shrublands (i.e., mostly native species) or semi-intact shrublands (i.e., native shrubs with understories dominated by non-native annuals) and grasslands (typically dominated by nonnative annuals, such as cheatgrass) (Fig. 1a–c). The corners of each plot were precisely located using a survey-grade GNSS receiver, and elevated reflector discs were deployed at each corner to provide control points for coregistration and georegistration of the TLS scans. The collection of TLS data was performed using a Riegl VZ-1000 near-infrared (1550 nm) scanner mounted on a 2-m tripod. The Riegl VZ-1000 is a time-of-flight waveform scanner, though the data were collected in discrete return mode. At a range of 100 m,


this instrument has a reported standard deviation error of 8 mm and a beam diameter of 30 mm (beam divergence is 0.3 mrad) (Riegl, Austria). Single-return scans were performed with 0.02 degrees of separation between pulses. Plots were scanned from five positions, once from the approximate midpoint of each side and once from the approximate plot center (Fig. 2). Slight leeway (< 5 m) in scanner location selection allowed for adaptation to reduce occlusion in each scan. The registration of the scans was performed in the RiSCAN Pro software package (Riegl, Horn, Austria). Using the reflectors in each scan, the coregistration resulted in a point cloud with approximately 5 cm absolute and 1.5 cm relative accuracy. Each plot was classified as shrub (n = 13) or grass (n = 13), based on the dominant community type (examples are shown in Fig. 1b, c). Plots that were classified as grass had experienced a severe standreplacing fire. In addition to cheatgrass and other herbaceous vegetation (generally < 25 cm tall), grass plots sometimes contained tall bunchgrasses and structurally sparse forbs. Vegetation in each plot was quantified manually at 9 systematically positioned points using several methods following Pilliod and Arkle (2013). Data collection included (1) a nadir photograph classified by land cover, allowing measurement of the unvegetated area over a 1 × 1.5 m extent; (2) the maximum height of vegetation falling in a 1-m2 quadrat

Fig. 2 The redundant sampling scheme used to test repeatability of point cloud datasets yielded by a 5-scan/plot methodology (n = 4). The green and gold squares represent differently aligned 1 ha plots, with green and gold circles representing their respective scan positions

Page 5 of 12 578

centered on each point was measured; (3) the distance in 4 cardinal directions from each point to the nearest tall bunchgrass and shrub was recorded, enabling estimation of the spatial density of individuals of both plant lifeforms; and (4) the canopy-spanning distance of the nearest tall bunchgrass and shrub in 4 cardinal directions from each point was measured along a line intercepting the point, allowing estimation of canopy cover fraction of both plant life-forms. The values recorded at all 9 points were averaged for each plot and reported for all grass and shrub plots. In 4 of the plots (shrub (n = 2) and grass (n = 2)), the scanning position layout was duplicated at a rotation of 45°, yielding 2 independent sets of 5 scan positions (Fig. 2). Although the square hectare plots aligned with each set of scans were rotationally offset, they shared a central 8283m 2 region and thus we used this area for our analyses. All TLS sampling was performed between 15 May and 14 June 2013. By this date, many non-native grasses and forbs, as well as some native species, were mostly senescent but structurally intact. Processing We performed our analyses in 2D and 3D, depending on the objective. We used pixels (2D representations of samples) to test the fraction of space that was sampled in different intervals of range from scanner or different layouts of scan positions. Because landscapes are continuous across a known area, we were able to evaluate the extent of the unsampled area (ground and vegetation with no points recorded in 2D) using pixels, whereas the unsampled volume occupied by aboveground matter (measured in 3D as voxels) is variable and unknown. Understanding the fraction of sampled area in 2D allows a practitioner to consider the quality and resolution of basic ecological measurements obtained, such as topography, vegetation height, and canopy openness. We tested the repeatability of a 5 scan position protocol using voxels (3D representations of samples) to evaluate the similarity of complete point clouds collected from offset scan position layouts. This 3D analysis helps to understand the extent to which fine-scale measurements of 3D vegetation structure are influenced by the specific geographic layout of scan positions.

578

Fig. 3 The center scan at each plot (n = 26) was divided into 3 m intervals from 3 to 6 m to 96–99 m from the scanner, and the fraction of sampled area in each interval was recorded. This figure shows a 1-ha shrub plot with light-colored sampled areas and black occlusion, with red bands representing the boundaries between range intervals

To test the effect of range in 2D on the fraction of unsampled area, the scan performed at the center of each plot was rasterized into binary pixel values. The binary classification indicated the presence or absence of one or more TLS returns (samples) measured within each pixel. The point density (beyond the presence or absence of samples in each pixel) was not considered. The number of pixels with samples falling in 3-m intervals in range from the scan position was tallied for the intervals 3–6 m through 96–99 m (Fig. 3). The fraction of sampled area in each interval ring was calculated (Eq. 1). This analysis was performed using 5, 20, and 50 cm pixels. These pixel sizes were chosen to roughly represent a range in 2D plant canopy size (i.e., small grasses to shrubs). Fraction of sampled area ¼


Page 6 of 12

number of sampled pixels pixel size ð1Þ area of the interval ring

To test the effectiveness of additional scanning positions at increasing the extent of sampling in datasets, we measured the sampled area yielded by 3 configurations of scan positions in each plot: a single center position, 3 in-line positions with 1 at the plot center and 2 on opposing edges, and 5 scans in a cross shape (Fig. 4). Each of these configurations exhibits fourfold symmetry

Fig. 4 The nested scan position configurations used to study the effect of additional scans. The configuration of one scan position uses just the central position, represented by a black star. The 2 sets of 3 scan positions are represented by the central position plus the edge positions represented by either the green or gold circles. The configuration of five scan positions uses all the positions symbolized

around the plot center. The configurations are nested in that the 3- and 5-scan configurations consist of the next smaller configuration plus 2 new scans. The point cloud corresponding to each configuration was rasterized into binary pixels indicating presence or absence of samples, and the number of sampled pixels was calculated for each pixel size using Eq. 2. Fraction of pixels sampled ¼

number of sampled pixels pixel size ð2Þ 1 ha

Since 2 arrangements of 3 in-line scan positions were possible with the available data, we used the average count of sampled pixels yielded by both. The analysis was performed for 5, 20, and 50 cm pixels. Additionally, this analysis was performed at the 4 plots where extra scanning was performed, allowing nested configurations of 1, 3, 5, and 9 scan positions which exhibited fourfold symmetry. Where multiple arrangements of each configuration existed, all were performed and the resultant counts of sampled pixels were averaged. We tested the reproducibility of sampling using differently positioned implementations of a geometrically consistent 5 scan position layout. We did this by comparing the independent point clouds yielded by the 2 sets of scans at the 4


Page 7 of 12 578

plots where redundant sampling (rotated by 45°) was performed (Fig. 2). Point clouds were cropped to the 8283-m2 region central to both sets of scans and generalized into voxels. For each plot, the fraction of voxels reproduced was calculated by Eq. 3. The use of Eq. 3 provides information on

Fraction of voxels reproduced ¼

the fraction of data which would be sampled reliably by the scanning position layout regardless of its particular alignment relative to the plot. The analysis was performed for 5, 10, 20, and 50 cm voxels.

number of voxels sampled commonly in both sets of scans number of voxels sampled inasingle scansetðaverage of the 2Þ

Results Field vegetation measurements The mean grass height and canopy cover was 44 cm and ~ 40% in grass-dominated plots. In shrub-dominated plots, the mean shrub height and canopy cover was 57 cm and 11%, respectively (Table 1). Elevation models derived from the TLS data indicated elevation gradients of < 5% in the plots. Effect of range on extent of occluded area Graphs showing the fraction of sampled area versus range from scanner using Eq. 3 are shown in Fig. 5. Except at very close ranges, the average fraction of sampled pixels was lower in shrub plots than grass plots at all pixel sizes. This corresponds to expectations that shrubs yield more extensive occlusion than grass. At longer ranges, the difference between the grass and shrub plots approaches 0 as fractional sampling

ð3Þ

coverage approaches 0 (most clearly demonstrated here using 5 cm pixels over the ranges considered; Fig. 5). As pixel size increased, a greater fraction of area was identified as sampled. This is expected as areas indicating sampling presence using small pixels will also indicate presence using larger pixels. As the pixel size used to quantify sampling coverage increases, the difference in mean sampling coverage between shrub and grass plots at a given range also increases. This is likely due to different mechanisms of occlusion. Large shadows produced by shrubs are detected using large pixel sizes, while the occlusion produced by grasses are often too narrow or discontinuous to be identified using large presence-absence pixel sizes, even if these shadows are spatially extensive. Fractional sampling coverage exhibits a plateau of high values (90–100%) at close ranges. The ranges where consistently high sampling coverages are achieved are longer in grass plots and when using larger pixel sizes for measurement. This effect is accounted for

Table 1 Vegetation metrics measured at each field plot aggregated by plot type (n = 13 of each) Vegetation metric

Grass Mean

Shrub Stdev

Mean

Stdev

Unvegetated area

19.9%

13.4%

33.5%

18.9%

Maximum height

44.3 cm

7.4 cm

56.5 cm

9.9 cm

Spatial density of tall bunchgrasses

0.2 m−2

0.2 m−2

0.1 m−2

0.1 m−2

Canopy cover of tall bunchgrasses

1.9%

2.8%

0.3%

0.4%

Spatial density of shrubs

0.1 m−2

0.3 m−2

0.9 m−2

0.8 m−2

Canopy cover of shrubs

1.6%

3.8%

11%

15.4%

Canopy cover of grasses and herbaceous vegetation

41.3%

34.9%

21.1%

28.2%

Stdev standard deviation

578

Page 8 of 12


threshold. A likely mechanism of this is that a protrusion will produce more occlusion as the line-of-sight vector from the scanner to the protrusion becomes more lateral. That the trend is more pronounced at smaller pixel sizes suggests that it is driven by occlusion caused by low stature vegetation, which are individually too small to be captured by large pixels. Sampling contribution of additional scan positions

Fig. 5 Fraction of sampled area in 3 m range intervals using 5, 20, and 50 cm pixel sizes to measure sampled area using Eq. 3. Gold lines show grass plots (n = 13) and green lines show shrub plots (n = 13), while bold and dashed lines show class means and ± 1 standard deviation, respectively

by the more-acute scan angles at close ranges, which prevents the occurrence of long lateral shadows. In grass plots, mean sampling coverage remained ≥ 90% up to ranges of 18 m using 5 cm pixels, 36 m using 20 cm pixels, and 66 m using 50 cm pixels. This sampling coverage among shrub plots was attained at ranges of 12 m using 5 cm pixels, 18 m using 20 cm pixels, and 30 m using 50 cm pixels. Measurements made with the smaller pixel sizes (5 and 20 cm) yield a relationship between range and fractional sampling coverage which resembles a negative logarithmic trend beyond the ~ 90% sampling

As discussed above, shrubs are expected to reduce measured sampling coverage by creating large occlusions, and using large pixels is expected to increase measured sampling coverage as small occluded areas are ignored. These assumptions were confirmed as we found that the mean area sampled in grass plots was higher than in shrub plots across configurations and pixel sizes and that there was a positive effect of increasing pixel size on measured sampling coverage yielded by each scan position configuration (Fig. 6). Due to the more complete coverage yielded by the initial scan in grass plots and analyses using large pixel sizes, there was less room for improvement with additional scans than in shrub plots and analyses using small pixel sizes. Likewise, the contribution of data by additional scans (Table 2) consistently decreases as the number of scans already performed increases. Additional scans increased the area sampled in all cases except in grass plots using a 50-cm pixel size, where 3 scans achieved complete sampling coverage. Sampling coverage averaging 99 or 100% was also achieved in grass with 5 scans using a 20-cm pixel size and in shrub plots with 5 scans using a 50-cm pixel size. The 4 plots where additional sampling was performed allow further inference about cases where complete sampling coverage could not be accomplished using the 5-scan configuration. For all 4 plots (2 grass and 2 shrub), nearly complete sampling (> 98%) was accomplished using 9 scans and a 20-cm pixel size. Using 9 scans and a 5-cm pixel size, 99% sampling coverage was accomplished for both grass plots, while shrub plots received sampling coverage of 80 and 93%. Reproducibility of 5-scan position methodology Reproduced sampling of voxels between repositioned deployments of a 5-scan sampling


Page 9 of 12 578

Fig. 6 The fraction of area sampled by increasing configurations of scan positions, measured using 5, 20, and 50 cm pixels and using Eq. 2. Gold lines show grass plots (n = 13) and green lines show shrub plots (n = 13), while bold and dashed lines show class

means and ± 1 standard deviation, respectively. At the left are values calculated for all plots, and at the right are values for the 2 grass plots and 2 shrub plots where additional sampling (up to 9 scans) was performed

methodology (see Fig. 4) ranged between about half to over 90% of voxels (Fig. 7), depending on the plot’s vegetation community and the voxel size used for comparison. Voxels sampled in grass plots were reproduced more reliably than voxels sampled in shrub plots. The rates of reproducibility more closely resembled one another between the two grass plots than the two shrub plots, reflecting the greater diversity in arrangement and composition of the

shrubs. A qualitative examination of the shrub plot datasets showed that voxels which were not reproduced tend to represent targets below the shrub canopy, mainly ground surface or low grasses. In the grass plots, the non-reproduced voxels were primarily of the ground surface that was occluded by grass and followed the pattern of grass distribution across the plots. The average percentage of voxels which occur in one set of scans and are reproduced in a

578


Page 10 of 12

Table 2 The mean percentage of area sampled by configurations of 1, 3, and 5 scan positions in 1 ha plots (n = 13 each of grass and shrub plots) Plot type Shrub

Grass

Pixel size

1 scan

3 scans

5 scans

5 cm

35%

55%

69%

20 cm

64%

82%

91%

50 cm

83%

95%

99%

5 cm

45%

68%

82%

20 cm

82%

95%

99%

50 cm

96%

100%

100%

second is 51% in shrub plots and 69% in grass plots using a 5-cm voxel size, 71% in shrub plots and 87% in grass plots using a 10-cm voxel size, 84% in shrub plots and 94% in grass plots using a 20-cm voxel size, and 93% in shrub plots and 97% in grass plots using a 50-cm voxel size.

Discussion A notable finding of this study is that low stature vegetation is an important mechanism of occlusion at long ranges. Although scanning was performed with a small enough angle between pulses to sample every pixel at a range of 99 m, at this range only 8% of 5 cm pixels, 37% of 20 cm pixels, and 68% of 50 cm pixels on average were sampled across grass plots using a single center scan. This shows that the extent of long-range sampling coverage in

Fig. 7 The fraction of voxels sampled in a single set of 5 scans which were also sampled in another set of scans from different positions (Eq. 3). Voxel sizes of 5, 10, 20, and 50 cm are considered. Gold lines show 2 grass plots and green lines show 2 shrub plots

rangelands is controlled not by the density of sampling, but by the height and density of vegetation, topographic complexity, and the height of the scanner. Similar results were found in Trochta et al. (2013), in which recognition of trees with single scans rapidly declined to 40–50% at distances of 35–40 m away. In an ecosystem similar to our study area, a methodology using several scan positions could be optimized by spacing scan positions at least twice the distance of the range of good performance (i.e., 90% sampling) at the scale of interest. This approach will prevent overly redundant collection. To better identify the optimal spacing between several scan positions, further research could consider the effect of range from two or more complementary positions on scanning coverage. We show that a study interested in grassland features at the 20-cm scale needs no more than 3 scan positions per hectare to provide complete sampling coverage, while a study focused on 5 cm scale features in shrublands (of similar height as our study area) needs greater than 9 scan positions per hectare to achieve complete sampling coverage. A limitation of these results is that they only apply to the specific scan configurations tested and our results for the 9-scan configuration is based on 2 plots. Wilkes et al. (2017) conclude that a 10 m × 10 m sampling grid is preferred regardless of forest type. While their study focused on forested sites, their suggestion of a continuous chain of 10 m × 10 m grid samples provide an upper and conservative threshold for sampling. Future work would be needed to discover trade-offs between various layouts of


a given number of scans. Overall, our work suggests the number of scans required to achieve a desired coverage is dependent on vegetation structure in the plot, which is congruent with Cifuentes et al. (2014). The results indicate that about 51 and 69% of sampled volume collected by 5 scans/ha in shrub and grass plots, respectively, are reproducible at 5 cm scales using the same sampling protocol, while sampling of the remaining fraction is restricted to specific alignments of the scan position layout. In contrast, nearly all of the sampled volume in both plot types was reproduced at the 50-cm scale. These results do not indicate overall thoroughness of sampling or any other description of dataset quality. However, our results may be used to help estimate the extent that TLS data are subject to topographic and vegetation effects associated with the scanner’s field-of-view, at least in ecosystems with vegetation cover and topographic complexity similar to our study area, and in which a similar layout of scan positions is used. These findings also strongly suggest that studies applying TLS to detect fine-scale vegetation change across large plots should take caution to use identical scanning positions at each sampling time, as using different sets of positions may falsely indicate a substantial amount of change over time. In change detection studies where positions are not reused exactly, our results may help to indicate potential error arising from inconsistency of scanning positions in the context of the landscape. For example, our results indicate that 2 scan collections using our protocol of 5 scans/ha where scan positions were not perfectly consistent could expect about 50% of 5 cm voxels to be unique to each collection, even if the landscape had not actually changed in the time between collections.

Conclusions The results presented here offer guidance to future researchers and practitioners planning the application of similar TLS methods and instruments to estimate 2D or 3D vegetation structural parameters (e.g., cover, biomass) across large plots with shrub and grass cover typical of rangeland systems. Understanding the potential of occlusion as a function of scan range, vegetation type, and scale of analysis, along with reproducibility will maximize efficiency of future TLS sampling. An important qualifier to our results and recommendations is that they may only apply to study areas with similar

Page 11 of 12 578

landscape characteristics to those we sampled. Our results are framed in the context of a study site that had relatively flat topography and sagebrush steppe vegetation cover typical of many dryland ecosystems. Different vegetation structure, such as taller or denser vegetation and more complex topography will increase occlusion and decrease reproducibility of the TLS collection methodologies. The spatial arrangement of vegetation across plots should also be considered. A priori understanding of the strengths and limitations of TLS sampling methods is vital to maximizing efficiency of sampling in logistically challenging field campaigns. The measures employed in each of our experiments changed greatly depending on the composition of vegetation in the plots and the scale at which measurements were made. By stratifying our results by both factors, we hope that future researchers and practitioners may identify the results most pertinent to their own work. With advance knowledge of study area composition, along with a predetermination of the resolution, coverage, and sampling reproducibility that is needed of datasets, our findings may be used to design efficient and logistically optimal TLS sampling protocols. We also encourage modification and expansion of our analytical approach, to continue to enhance the utility of TLS data collection under a variety of environmental conditions. Acknowledgements This work was supported by a Joint Fire Science Program grant (Project ID: 11-1-2-30), National Oceanic and Atmospheric Administration’s Earth System Research Labor a t o r y ( E S R L , P h y s i c a l S c i e n c e s D i v i s i o n ) Aw a r d NA10OAR4680240, the National Science Foundation Idaho Experimental Program to Stimulate Competitive Research Program, and the NSF under award number EPS-0814387. We thank Randy Lee at Idaho National Laboratory for the use of the TLS; Dr. Rupesh Shrestha, Dr. Aihua Li, Mr. Peter Olsoy, Mr. Kyle Gochnour, and Mr. Samuel Gould for providing lab and field assistance; and Dr. Steve DeLong for providing a constructive review. Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the US Government. This article has been peer reviewed and approved for publication consistent with US Geological Survey Fundamental Science Practices (http://pubs.usgs.gov/circ/1367).

References Andersen, H. E., McGaughey, R. J., & Reutebuch, S. E. (2005). Estimating forest canopy fuel parameters using lidar data. Remote Sensing of Environment, 94(4), 441–449. Anderson, K. E., Glenn, N. F., Spaete, L. P., Shinneman, D. J., Pilliod, D. S., Arkle, R. S., McIlroy, S. K., Derryberry, D. R. (2018). Estimating vegetation biomass and cover across large

578

Page 12 of 12

plots in shrub and grass dominated drylands using terrestrial lidar and machine learning. Ecological Indicators, 84:793802 Brooks, M. L., D’antonio, C. M., Richardson, D. M., Grace, J. B., Keeley, J. E., DiTomaso, J. M., Hobbs, R. J., Pellant, M., & Pyke, D. (2004). Effects of invasive alien plants on fire regimes. Bioscience, 54(7), 677–688. Chasmer, L., Hopkinson, C., & Treitz, P. (2004). Assessing the three-dimensional frequency distribution of airborne and ground-based lidar data for red pine and mixed deciduous forest plots. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 36, 8. Cifuentes, R., Van Der Zande, D., Farifteh, J., Salas, C., & Coppin, P. (2014). Effects of voxel size and sampling setup on the estimation of forest canopy gap fraction from terrestrial laser scanning data. Agricultural and Forest Meteorology, 194, 230–240. Clawges, R., Vierling, L., Calhoon, M., & Toomey, M. (2007). Use of a ground-based scanning lidar for estimation of biophysical properties of western larch (Larix occidentalis). International Journal of Remote Sensing, 28(19), 4331–4344. Cooper, S. D., Roy, D. P., Schaaf, C. B., & Paynter, I. (2017). Examination of the potential of terrestrial laser scanning and structure-from-motion photogrammetry for rapid nondestructive field measurement of grass biomass. Remote Sensing, 9(6), 531. Coveney, S., & Fotheringham, A. S. (2011). Terrestrial laser scan error in the presence of dense ground vegetation. The Photogrammetric Record, 26(135), 307–324. Fan, L., Powrie, W., Smethurst, J., Atkinson, P. M., & Einstein, H. (2014). The effect of short ground vegetation on terrestrial laser scans at a local scale. ISPRS Journal of Photogrammetry and Remote Sensing, 95, 42–52. Greaves, H. E., Vierling, L. A., Eitel, J. U. H., Boelman, N. T., Magney, T. S., Prager, C. M., & Griffin, K. L. (2015). Estimating aboveground biomass and leaf area of lowstature arctic shrubs with terrestrial lidar. Remote Sensing of Environment., 164, 26–35. Henning, J. G., & Radtke, P. J. (2006). Ground-based laser imaging for assessing three-dimensional forest canopy structure. Photogrammetric Engineering & Remote Sensing, 72(12), 1349–1358. Heritage, G., & Large, A. (2009). Laser scanning for the environmental sciences. Hoboken: John Wiley & Sons. Loudermilk, E. L., Hiers, J. K., O’Brien, J. J., Mitchell, R. J., Singhania, A., Fernandez, J. C., Cropper, W. P., & Slatton, K. C. (2009). Ground-based lidar: a novel approach to quantify fine-scale fuelbed characteristics. International Journal of Wildland Fire, 18(6), 676–685. Murgoitio, J., Shrestha, R., Glenn, N., & Spaete, L. (2014). Airborne lidar and terrestrial laser scanning derived vegetation obstruction factors for visibility models. Transactions in GIS, 18(1), 147–160. Olsoy, P. J., Glenn, N. F., Clark, P. E., & Derryberry, D. R. (2014). Aboveground total and green biomass of dryland shrub

Environ Monit Assess (2017) 189:578 derived from terrestrial laser scanning. ISPRS Journal of Photogrammetry and Remote Sensing, 88, 166–173. Olsoy, P. J., Mitchell, J. J., Levia, D. F., Clark, P. E., & Glenn, N. F. (2016). Estimation of big sagebrush leaf area index with terrestrial laser scanning. Ecological Indicators, 61, 815–821. Pilliod, D. S., & Arkle, R. S. (2013). Performance of quantitative vegetation sampling methods across gradients of cover in Great Basin plant communities. Rangeland Ecology and Management, 66, 634–647. Rowell, E., Loudermilk, E. L., Seielstad, C., & O’Brien, J. J. (2016). Using simulated 3D surface fuelbeds and terrestrial laser scan data to develop inputs to fire behavior models. Canadian Journal of Remote Sensing, 42(5), 443–459. Seidel, D., Fleck, S., & Leuschner, C. (2012). Analyzing forest canopies with ground-based laser scanning: a comparison with hemispherical photography. Agricultural and Forest Meteorology, 154, 1–8. Shan, J., & Toth, C. K. (2008). Topographic laser ranging and scanning: principles and processing. CRC Press. Shinneman, D., Arkle, R., Pilliod, D., Glenn, N. F. (2015). Quantifying and predicting fuels and the effects of reduction treatments along successional and invasion gradients in sagebrush habitats. Retrieved from: https://www.firescience. gov/projects/11-1-2-30/project/11-1-2-30_final_report.pdf Strahler, A. H., Jupp, D. L., Woodcock, C. E., Schaaf, C. B., Yao, T., Zhao, F., Yang, X., Lovell, J., Culvenor, D., Newnham, G., Ni-Miester, W., & Boykin-Morris, W. (2008). Retrieval of forest structural parameters using a ground-based lidar instrument (Echidna®). Canadian Journal of Remote Sensing, 34(2), 426–440. Trochta, J., Král, K., Janík, D., & Adam, D. (2013). Arrangement of terrestrial laser scanner positions for area-wide stem mapping of natural forests. Canadian Journal of Forest Research, 43, 355–363. USDI Bureau of Land Management. (2008). Snake river birds of prey National Conservation Area proposed resource management plan and final environmental impact statement, Boise, Idaho. Van der Zande, D., Jonckheere, I., Stuckens, J., Verstraeten, W. W., & Coppin, P. (2008). Sampling design of ground-based lidar measurements of forest canopy structure and its effect on shadowing. Canadian Journal of Remote Sensing, 34(6), 526–538. Vierling, L. A., Xu, Y., Eitel, J. U., & Oldow, J. S. (2013). Shrub characterization using terrestrial laser scanning and implications for airborne lidar assessment. Canadian Journal of Remote Sensing, 38(6), 709–722. Wallace, L., Lucieer, A., Watson, C., & Turner, D. (2012). Development of a UAV-lidar system with application to forest inventory. Remote Sensing, 4(6), 1519–1543. Western Region Climate Center. 2012. Swan falls power house, Idaho, Period of record general climate summary. Retrieved from: http://www.wrcc.dri.edu/cgi-bin/cliMAIN.pl?id8928. Wilkes, P., Lau, A., Disney, M., Calders, K., Burt, A., de Tanago, J. G., Bartholomeus, H., Brede, B., & Herold, M. (2017). Data acquisition considerations for terrestrial laser scanning of forest plots. Remote Sensing of Environment, 196, 140–153.