- Email: [email protected]
Assessing urban tree condition using airborne light detection and ranging
Accepted Manuscript Title: Assessing urban tree condition using airborne light detection and ranging Author: Andrew A. Plowright Nicholas C. Coops Bianca N.I. Eskelson Stephen R.J. Sheppard Neal W. Aven PII: DOI: Reference:
S1618-8667(15)30064-9 http://dx.doi.org/doi:10.1016/j.ufug.2016.06.026 UFUG 25736
To appear in: Received date: Revised date: Accepted date:
7-10-2015 12-5-2016 30-6-2016
Please cite this article as: Plowright, Andrew A., Coops, Nicholas C., Eskelson, Bianca N.I., Sheppard, Stephen R.J., Aven, Neal W., Assessing urban tree condition using airborne light detection and ranging.Urban Forestry and Urban Greening http://dx.doi.org/10.1016/j.ufug.2016.06.026 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Assessing urban tree condition using airborne light detection and ranging Andrew A. Plowrighta M.Sc. candidate [email protected] * Corresponding author
Nicholas C. Coopsa Professor [email protected]
Bianca N.I. Eskelsona Assistant professor [email protected]
Stephen R.J. Shepparda Professor [email protected]
Neal W. Avenb Urban Forestry & Environmental Programs Manager [email protected]
a
University of British Columbia
Department of Forest Resources Management Faculty of Forestry Vancouver, Canada
b
City of Surrey
1
Urban Forestry and Environmental Programs Parks Division Surrey, Canada
Permanent address: IRSS: Integrated Remote Sensing Studio Department of Forest Resources Management Room 2231 2424 Main Mall Vancouver, BC, Canada V6T 1Z4
Highlights
Airborne LiDAR was used to estimate two indicators of urban tree condition: tree height and crown density. An automated method for detecting urban trees supported by GIS data was developed and tested. The coefficient of variation of LiDAR return heights is a stronger predictor of crown density than the percentage of non-ground returns. Improved methods for estimating crown density for juvenile trees are needed.
Abstract With increased interest in urban forests on behalf of city dwellers and urban planners, there is a growing need for comprehensive information on urban tree condition. This study examines the potential of airborne light detection and ranging (LiDAR) for evaluating tree condition in the urban center of Surrey, Canada. An approach to detecting and outlining free-growing trees from LiDAR data augmented by a municipal tree inventory was developed and validated. Once the trees were located, LiDAR was used to 2
estimate two field-measured indicators of tree condition: crown density and tree height. Tree heights estimated by LiDAR were, as expected, well correlated with field measurements (Pearson’s r = 0.927, p < 0.001), indicating accurate height estimates of successfully detected trees. Two LiDAR metrics, the percentage of non-ground LiDAR returns and the coefficient of variation of return height, were examined as predictors of crown density. While the percentage of non-ground returns performed relatively poorly (r2 between 0.005 and 0.23 across multiple tree height classes), the coefficient of variation of return height was able to predict crown density with an r2 = 0.617 for trees over 8 m. In addition, residuals derived using expected height growth from the known planting date and their LiDARderived height was found to be a useful tree condition metric. We conclude that despite the complexity of urban tree condition assessment, airborne LiDAR is a promising tool for detecting trees in an urban environment and measuring indicators of their condition.
Key words: LiDAR, GIS, remote sensing, urban trees, tree condition, tree detection
1. Introduction 1.1 The need for urban tree condition assessment Through their numerous aesthetic, environmental and economic benefits, trees are a key component to sustainable urban development. As the scientific foundation of urban forestry matures and the popular demand for green space increases, many city governments have undertaken ambitious projects to 3
expand their tree cover (City of Toronto, 2008; Locke et al., 2010; McPherson, Simpson, Xiao, & Wu, 2011). However, trees in urban environments are exposed to a wide variety of stress-inducing agents, such as unsuitable soil conditions, rerouted water flows, mechanical damage, and pollution (Jim, 1993, 1998; Malthus & Younger, 2000). As a result, urban forests are intensively managed, with expenditures towards urban trees per acre in the United States surpassing those in forested lands outside of cities (McPherson, 1993). With the considerable financial investment represented by urban forests, it is incumbent on city authorities to assess the health and vitality of the trees under their purview. Although hypothetical optimal tree vitality is difficult to define, the effects of environmental stressors are reflected by a range of clearly identifiable biochemical, physiological and structural symptoms (Dobbertin, 2005). The overall condition of a tree can be defined as the culmination of these symptoms (Stone, Coops, & Culvenor, 2000). Citywide data on tree condition can help optimize the allocation of resources and guide effective management prescriptions, such as targeted irrigation, pruning or removal, yet city authorities rarely have access to this type of information. Exhaustive ground surveys are prohibitively expensive, while sampling-based approaches may not be appropriate due to the wide range of tree species, age classes and the spatial heterogeneity of the urban environment (McPherson, 1993). 1.2 The use of remote sensing for assessing tree condition By providing efficient and repeatable means of acquiring quantitative and spatially explicit data, remote sensing may have the potential to address the need for comprehensive urban tree condition information. Given their capacity to detect changes in the reflectance of tree foliage, optical remote sensors have received considerable attention for this purpose (Goodwin, Coops, & Stone, 2005). Manifestations of tree stress, such as leaf chlorosis or necrosis, affect leaf spectral reflectance. Airborne multispectral imagery has been used to estimate chlorophyll content as an indicator of forest health
4
(Sampson, Zarco-Tejada, Mohammed, Miller, & Noland, 2003), detect root disease (Reich & Price, 1999), and assess tree condition in urban areas (Malthus & Younger, 2000; Nowak & McBride, 1991). Optical sensors are limited, however, when attempting to characterize a tree’s vertical structure. The dimensions and the vertical architecture of a tree are reflective of its net primary productivity and as such, are key indicators of its overall health and vigor (Schomaker et al., 2007). A large, dense crown is an indicator of optimal tree growth, while sparsely foliated crowns are a sign of deterioration and stress (Zarnoch, Bechtold, & Stolte, 2004). Tree height can also be used to assess a tree’s condition, as healthy trees grow rapidly while stressed trees experience stunted growth (Dobbertin, 2005). Although some vertical tree metrics can be estimated through indirect relationships with combinations of spectral bands (Cohen & Spies, 1992), airborne light detection and ranging (LiDAR) may provide the means to measure them directly. A LiDAR instrument emits pulses of light that are reflected off trees, ground surfaces, and other terrestrial features. Of particular note is LiDAR’s capacity to penetrate through gaps in the foliage, enabling it to directly measure the vertical aspects of tree crowns and forest canopies (Coops et al., 2007) While extensive research has demonstrated the utility of LiDAR in natural resource disciplines such as forestry, ecology, wildlife management and hydrology, its uses in urban forestry remain nascent. Zhang, Zhou, & Qiu (2015) developed an automated algorithm for detecting, outlining and mensurating urban trees directly from a LiDAR point cloud. Several studies have focused on integrating LiDAR with spectral imagery to map urban tree canopy cover (MacFaden, O’Neil-Dunne, Royar, Lu, & Rundle, 2012; O’NeilDunne, MacFaden, Royar, & Pelletier, 2013). LiDAR data can be particularly useful in dense urban areas where the shadowing effect of buildings make the use of spectral imagery problematic (MacFaden et al., 2012). Other applications of LiDAR and high resolution imagery include mapping urban tree species (C. Zhang & Qiu, 2012), modeling solar radiation effects (Tooke, Coops, & Voogt, 2009), estimating citywide
5
carbon storage (Schreyer, Tigges, Lakes, & Churkina, 2014) and detecting invasive plant species (Singh, Davis, & Meentemeyer, 2015). 1.3 Research objectives This paper explores the potential of airborne LiDAR as a tool for assessing urban tree condition within the city of Surrey, Canada. The objectives of this study are to: 1. Develop and evaluate an automated method for locating and outlining urban trees that is supported by ancillary GIS data collected by the city. 2. Examine LiDAR’s capacity to estimate two dendrological metrics: tree height and crown density. Methods for using these metrics as indicators of urban tree condition are presented and discussed. Our analysis focuses on western redcedar trees (Thuja plicata) located on city property. Western redcedar is the official tree of the Canadian province of British Columbia, and possesses unique cultural significance in the region. Furthermore, the species is of particular concern to city managers due to its preference for moist soils, and consequently its susceptibility to poor watering conditions (Stewart, 1984).
2. Data & study site 2.1 Study site The city of Surrey is located in the Greater Vancouver regional district, in the province of British Columbia, Canada. It is one of the fastest growing cities in Canada, with an 18.6% increase in population between 2006 and 2011 (Statistics Canada, 2011). Over 90,000 trees are actively managed on city property, with 3,500 to 5,000 additional trees being planted every year. As part of its annual tree maintenance budget, the city spends roughly 600,000 USD on watering alone. The current methods 6
employed for monitoring Surrey’s trees are mainly field-based, such as performing soil moisture spot checks in the summer drought months. 2.2 GIS database The city of Surrey maintains a comprehensive GIS database of all the trees that it plants and manages. Each entry includes a tree’s species, subspecies, planting date and approximate geographic coordinates. Tree coordinates are either recorded by field crews using GPS receivers, or, for trees predating the database’s creation, located through aerial photo interpretation. Due to inaccuracies inherent to mobile GPS units and the variable methods used for recording coordinates, certain entries contain positional errors, while others may be out-of-date or supply locations for trees that have been removed. 2.3 LiDAR data and derived products Airborne LiDAR data was acquired over Surrey by Airborne Imaging (Calgary, Alberta), under contract with the city in April, 2013. Trees were under leaf-off conditions. A Leica ALS70-HP discrete return LiDAR system, with up to four discrete returns per pulse, was flown at 1000 m above ground level with 688 m swaths with 50% overlap. The pulse rate was 500 KHz, which resulted in an average first-return density of 25 points per square meter. Before being delivered by the contractor, the raw LiDAR point cloud was classified into land class covers, such as ground, building or vegetation, using TerraScan software (TerraSolid Ltd., Helsinki). A 1 m2 rasterized digital elevation model (DEM) interpolated from classified ground points using a triangular irregular network (TIN) was also supplied.
7
3. Methods 3.1 Field data A ground survey was undertaken of 169 western redcedar trees, all of which had corresponding entries in the city’s database. The focus of this study was on detecting trees whose growth is affected by environmental stressors as opposed to competition with neighboring trees, and so the sample was restricted to free-growing trees. The criteria for defining a free-growing tree was that its crown not be in contact with the crown of any trees in its surroundings. Trees planted in parks, parking lots and along roadsides were included. In March, 2015, an initial four sampling sites were visited. For logistical reasons, only sites with high numbers of western redcedar were considered. Accompanying a surge in urban development, a substantial portion of the city’s western redcedar were planted in the late 1990s and early 2000s, and consequently, the four initial sites contained mostly trees planted within this period. To ensure that the sample represented a sufficiently broad range of ages, three additional sampling sites were subsequently visited, containing trees under 15 years and over 25 years. The combined sample from the seven sites contained 169 free-growing western redcedars with ages ranging from 11 to 33 years old (Figure 1). Surrey’s urban trees are approximately seven years old when planted. Tree heights were measured using a LTI TruPulse 360 laser rangefinder. Crown density was measured using methods described in Schomaker et al. (2007). Crown densities ranged from 25% to 95%, while heights ranged from 2.1 m to 14.5 m. 3.2 Canopy Height Model A canopy height model (CHM) was produced using FUSION software (McGaughey, 2014). This product is derived from the raw LiDAR data by recording the difference in elevation between the highest classified
8
vegetation return and the underlying DEM within each cell of a continuous grid. The maximum aboveground height of vegetation is indicated by the pixel value of the CHM at any given location. The selection of an appropriate cell size is an important parameter: large cells reduce variation within the CHM, while small cells can create gaps within the canopy and increase the volume of data. Chen, Baldocchi, Gong, & Kelly (2006) recommend calculating an optimal cell size based on the first-return density of the LiDAR data (Equation 1), where 𝜆 is returns/m2, and 𝑑 is cell size.
d
1
(1)
Using a density of 25 returns/m2, this equation yielded a cell size of 0.2 m. In practice, however, return density is not uniform, and so a conservative cell size of 0.5 m was selected to account for areas where flight swaths did not overlap. By computing the return count in areas free of water, it was found that approximately 99.94% of cells contained at least one return when using this resolution. 3.3 Locating trees For information on tree condition to be extracted from the LiDAR data, the location and spatial extent of the trees of interest are required. A substantial amount of research has been conducted on methods of locating and outlining individual trees from LiDAR data. Automated tree detection algorithms can either be applied directly to a point cloud or to a rasterized canopy height model (CHM) derived from raw LiDAR data (Jakubowski, Li, Guo, & Kelly, 2013). Approaches such as valley following and object-oriented classification have been applied to coregistered LiDAR data and multispectral imagery to detect and isolate individual trees in coniferous forests (Leckie et al., 2003; Suárez, Ontiveros, Smith, & Snape, 2005). Kaartinen et al. (2012) compared a variety of approaches of automated tree detection, and found that several outperformed manual processing of LiDAR data in terms of accuracy.
9
This study employed the popular and well-researched variable window filter (VWF) method for detecting trees (Popescu, Wynne, & Nelson, 2002; Popescu & Wynne, 2004). A VWF uses a moving focal window to scan the canopy for local canopy height maxima, which often correspond to treetops. When applied to a CHM, cells are tagged as treetops if their values are the highest within the window. Errors of commission occur when the window size is too small, causing branches, canopy protrusions, or other non-tree local maxima to be incorrectly tagged as treetops. Conversely, an overly large window may pass over treetops that are closely clustered together, causing errors of omission. A VWF minimizes both types of error by adjusting the size of the window according to the height of the cell on which it is centered. This is based on the relationship between tree height and crown width: pixels with high values are likely to belong to tall trees with correspondingly wide crowns, and vice versa. To parameterize the VWF, a formula relating tree height and crown width is required. Crown outlines for a randomly selected subset of 115 surveyed trees were manually delineated through visual interpretation of the CHM. Using a quadratic model to regress the average crown diameter of these reference outlines with their height produced Equation 2, which yielded an adjusted r2 value of 0.5958. ̂𝑊𝑖𝑛𝑑𝑜𝑤 when centered over The equation is then used to derive the diameter of the search window 𝐷 any given CHM cell with a height value of HCell. 2 Dˆ W in d o w 0 .6 9 9 0 6 0 .6 6 5 1 1 H C e ll 0 .0 1 0 5 2 H C e ll
(2)
While Surrey’s GIS database contains the approximate locations of all the trees maintained by the city, coordinates may be offset from the trees’ actual locations, and some trees may be missing entirely. The following process (Figure 2) uses a VWF to complement the city’s GIS data and locate trees with increased precision. A 20 m2 square search area, approximately the size of the largest possible tree crown for western redcedar, is centered on a tree’s coordinates as recorded in the database. To prevent the detection of 10
false treetops from shrubs or other low-lying non-tree objects, all CHM pixels with a height value less than 1 m are masked out within the search area. Then, the VWF is applied to locate all potential treetops within the tree’s immediate vicinity. The potential treetops identified by the VWF are then filtered according to height. Commonly used growth curves for coastal western redcedar developed by Kurucz (1985) provide estimates of a tree’s height over time for any given site index. Site indices, which are expressed as the height of a dominant tree in meters at age 50, measure the growing potential of a forested site, with an elevated site index indicating optimal growing conditions and vice versa (Clutter, Fortson, Pienaar, Brister, & Bailey, 1983). Site indices of 40 and 20 were selected as upper and lower thresholds, which correspond to the range found within the field sample and are considered to represent the range of growing conditions found in the city. Using Site Tools software (BC Ministry of Forests, 2014), maximum and minimum values HTreetopMax and HTreetopMin were computed for each tree according to its age. Treetops outside of these thresholds were removed, and the remaining treetop closest to the tree’s coordinates was then selected. If no treetops within these thresholds are found, the tree is reported as missing. 3.4 Outlining tree crowns Once the coordinates of a given tree have been located, an outline of its crown’s horizontal extent is needed. An image processing technique known as watershed segmentation, originally developed for delineating drainage basins (Beucher & Lantejoul, 1979), can be used to segment tree crowns due to the morphological similarity between canopies and topographical terrain models (Figure 3). To avoid oversegmentation, wherein branches or other protrusions in the canopy are assigned spurious segments, the process is constrained by the pre-defined locations of treetops as detected by the VWF (Chen et al., 2006; Schardt, Ziegler, Wimmer, & Wack, 2002; J. Zhang, Sohn, & Brédif, 2014).
11
A second automated outlining approach, wherein the crown’s horizontal size and shape are approximated using a circle, was also tested. The average crown width of each tree is computed in a process described by McGaughey (2014). This approach determines the crown radius along 16 equally spaced radial profiles centered on a tree’s apex. Along each profile, the radius is given by the distance between the apex and the first CHM pixel that is either A) a local minima within the profile or B) less than 50% of the tree’s height. A circular outline is then generated using this average of these radii. Both automated outlining approaches were compared. Brandtberg et al. (2003) developed a fuzzy method for assessing the similarity between two corresponding tree crown outlines. This approach, which uses spatial weighting to assign greater importance to the center of an outline than its periphery, produces a segmentation assessment value (A) between two outlines, with an A = 1 indicating a perfect match, and any other case an A < 1. The ratio between the areas of automated outlines and their corresponding reference outlines (Q) was also computed. 3.5 Estimating crown density In a technique similar to the use of a cookie-cutter, LiDAR point cloud subsets corresponding to individual trees were created according to the boundaries of the crown outlines. The vertical measurements of the points in these subset were then normalized according to the DEM. Two metrics were tested as potential predictors of crown density. Previous attempts at estimating the foliar density of forests at an area-based level have used the ratio of LiDAR returns below a certain height threshold over the total number of returns (Lovell, Jupp, Culvenor, & Coops, 2003). The reasoning behind this metric is that denser canopies are more likely to intercept LiDAR pulses, preventing a larger portion of them from reaching the ground. Here, the percentage of returns with a height above ground higher than 0.5 m within each tree’s normalized LiDAR point cloud subset will be referred to as “percentage of non-ground returns”. 12
Other studies have used the coefficient of variation (CV) of the height of LiDAR returns to estimate forest attributes (Næsset, 2002). A large CV indicates a wide dispersion of LiDAR returns along the vertical profile, suggesting increased penetration of the canopy. While this may hold true for an area of continuous forest, certain considerations should be made for an individual tree. For instance, in an open area where the laser path is unobstructed by neighboring objects, the conical shape of the western redcedar may intercept LiDAR pulses evenly along the vertical profile, regardless of the crown’s density (Figure 4). To correct for this, the size of the outline used to subset the LiDAR points was halved, so that a large portion of the returns from the crown’s exterior are excluded. The coefficient of variation of the height of the returns within this subset will be referred to as “CV of return heights”. To remove the potential confounding effects of tree height, the reference trees were separated into 3 m equal interval height classes spanning the range of tree heights as measured in the field (Table 1). Due to the small number of trees in the highest class (14 trees between 11 m and 14 m) it was merged with the preceding class to ensure that all classes had a minimum of 30 trees. 3.6 Tree height as an indicator of condition The highest point within each LiDAR point subset was used as an estimate of the tree’s height. For a tree’s height to be an informative measure of its condition, it needs to be compared to a reference of growth for that tree’s species. With the availability of planting dates from Surrey’s GIS database, it is possible to fit growth models to trees whose heights have been accurately measured using LiDAR. The residual of a tree’s height against this model can then be used as an indicator of the tree’s condition (Dobbertin, 2005). Trees with positive residuals are growing at an optimum rate, while trees with negative residuals may be experiencing some sort of stress that is impeding growth. Based on methods described in Fekedulegn et al. (1999), a Chapman-Richards model was fit to the field-measured trees using the minpack.lm package for R statistical software (Elzhov, Mullen, Spiess, & Bolker, 2015). The
13
model’s general formula is given by Equation 3, where Hi is the height of a given tree (i), agei is its age according to the city’s database, εi is the model’s residual and 𝛽 are estimated coefficients:
H i 1 1 e
2 a g ei
3
i
(3)
4. Results 4.1 Treetop location and tree height estimation Of the 169 reference trees measured in the field, the automated tree detection algorithm reported one as missing. Figure 5 shows the relationship between tree heights measured in the field and heights measured from the LiDAR data, which had a Pearson correlation coefficient of r = 0.927, p < 0.001. Trees for which the difference between LiDAR- and field-based height measurements was more than three standard deviations from the mean were selected for visual inspection. The inspection revealed that these four outliers had been incorrectly identified by the algorithm and were removed. The remaining LiDAR heights were, on average, 0.72 m lower than those measured in the field, with a standard deviation of 1.09 m. Figure 6 shows the Chapman-Richards growth model fitted to the 169 field trees using their LiDAR-based height measurements. 4.2 Tree crown outlines On average, the ratio between the reference and the watershed segmentation outline areas (Q) was Q = 1.06, and Q = 1.01 for the circular outlines. Manually delineated reference outlines produced a segmentation assessment value (A) of A = 0.71 when compared to the watershed segmentation outlines and A = 0.63 when compared to the circular outlines. Figure 7 shows examples of various segmentation assessment values A. The standard deviation of A for watershed segmentation outlines was 0.21, with 19% of the outlines being nearly identical to their references (A > 0.9), and 23% of the outlines being
14
substantially different (A < 0.5). In comparison, the standard deviation for circular outlines was 0.14, with no outlines at A > 0.9, but only 17% with A < 0.5. 4.3 Crown density estimation Figure 8 shows ordinary least-squares linear regression models with field-measured crown density as the response variable and the percentage of non-ground LiDAR returns and the CV of return heights as separate predictor variables. For each of the three height classes, CV of return heights was a better predictor based on r2 values. Both predictor variables performed best in the highest height class (8 m to 14 m) and poorest in the smallest height class (2 m to 5 m). With an r2 = 0.617, predictions were best for the 8 m to 14 m height class using the CV of return heights. In order to demonstrate how the estimates of tree condition could be applied over the urban landscape, Figure 9 shows a comparison between tree height residuals and crown density for 40 of the field-measured trees in a subset of the residential area of Surrey. The map shows some apparent clumping in the spatial arrangement of tree condition indicators. Investigation of the drivers of these spatial patterns is the logical subsequent analysis which will follow in future work.
5. Discussion 5.1 Accuracy of treetop location algorithm For an automated algorithm to accurately estimate crown density from LiDAR data at an individual tree basis, the precise location of that tree is required. It should not be expected that city GIS databases, when available, will contain tree coordinates that meet this level of precision. Therefore, it is important to incorporate steps that will account for this potential inaccuracy, both by correcting positional errors and by skipping trees that may have been removed or incorrectly recorded.
15
Here, a VWF was applied to identify all potential treetops in the vicinity of a tree’s entry in the city’s GIS database. When used in forestry research, tree extraction routines such as the VWF are generally applied over large areas using a single set of parameters. These parameters are calibrated according to the forest’s species and age composition. This task is challenging in urban forests, however, whose highly variable structure make parameterizing automated algorithms difficult. Surrey’s existing GIS data can be used to address this issue. Although we focus here on a single species (Thuja plicata), speciesspecific equations for deriving search window sizes (such as Equation 2) can be developed and applied when extending this approach to other types of trees. Once a series of potential treetops have been identified, the treetop matching the tree entry must be selected. Using growth curves to compute a range of likely heights based on the tree’s age, and then using this range to filter potential treetops identified by the VWF was found to be a reliable way of locating trees within the LiDAR data and avoiding the incorrect selection of neighboring trees or false treetops such as buildings or power lines. The efficacy of this method was demonstrated by the highly significant correlation between height estimates obtained from the LiDAR data and heights measured in the field. 5.2 Performance of crown outline methods In comparison with previous methodologically similar studies (Brandtberg, Warner, Landenberger, & McGraw, 2003), both automated approaches for generating tree crown outlines produced acceptable results. The average ratio between the area of the circular outlines and their corresponding references (Q = 1.02) suggests this is an accurate method for measuring crown diameter. This approach produced a lower average segmentation assessment value (A = 0.63) than the watershed segmentation algorithm (A = 0.71), which is likely due to a mismatch between the circles and the sinuated contour of the tree’s fringe. It should be noted that western redcedars generally have a circular shape, and that the quality of
16
a circular outline will likely be reduced for trees with irregularly shaped crowns, and particularly those with off-center apexes. The standard deviation (SD) of A was highest for watershed segmentation outlines (SD = 0.21) which indicates a wider variation in the quality of the outlines. This is explained by this method’s reliance on accurate prior treetop detection. When all treetops are successfully detected within the vicinity of the targeted tree, watershed segmentation has the potential to produce highly accurate outlines, as demonstrated by the 19% of outlines with an A > 0.9. Conversely, watershed segmentation can oversegment a single tree when spurious treetops are mistakenly identified within its crown, or include multiple tree crowns in a single outline when treetops are missed. Although the watershed segmentation algorithm outperformed the circular outlines with respect to the average A, its higher incidence of unreliable outlines (23% of watershed segmentation outlines with A < 0.5, compared to 17% for circular outlines), has important implications when attempting to extract LiDAR metrics. For instance, an outline that erroneously includes neighboring trees will result in a highly inaccurate LiDAR-based estimate of crown density. Furthermore, Figure 4 illustrates the rationale for using only the horizontal center of a tree’s crown for density estimation. For this reason, the circular outlines were used to estimate crown density in lieu of those generated using watershed segmentation. 5.3 Potential use of LiDAR metrics for assessing urban tree condition While previous studies have used the ratio of LiDAR returns intercepted over a certain height (usually between 0.5 m to 2 m) to measure foliar density over large areas, this method does not appear to be applicable at the individual tree level. The percentage of returns above 0.5 m was a poor predictor of crown density for all tree height classes. Potential reasons include varying live crown ratios (the ratio of the vertical length of the crown to the tree’s full height) or interference from surrounding objects.
17
Figure 10 illustrates how neighboring trees or buildings could potentially affect the number of returns to reach the ground under a tree. The coefficient of variation of LiDAR return heights is a stronger predictor of density, particularly for trees in the 8 m to 14 m height class (r2 = 0.617). Taller trees intercept a greater number of LiDAR returns, which may reduce variability and improve the relationship between the two variables. According to estimates made by applying the treetop detection algorithm citywide, approximately 69% of Surrey’s urban western redcedars are taller than 8 m, suggesting that the coefficient of variation of return heights may be a useful measure of crown density for a large portion of Surrey’s urban trees. It should be noted, however, that the coefficient of variation is influenced by the number of LiDAR returns per tree and so if this metrics is applied in future work, transformations that account for variable LiDAR return density levels should be applied. Previous studies have established LiDAR as a reliable tool for measuring individual tree heights (Hyyppä & Inkinen, 1999). LiDAR measurements tend to underestimate tree height, however, as pulses have little chance of hitting a tree’s apex. Tree growth occurring between LiDAR and field measurements may have also contributed to this effect. These underestimations are generally consistent, as they are here by an average of 0.72 m, which allows them to potentially be corrected using empirically-derived offsets (Schardt et al., 2002). The correlation between tree heights estimated by LiDAR and heights measured in the field demonstrate both the functionality of the treetop detection method and the accuracy of LiDAR height estimates. By extension, these accurate height estimates support this approach’s potential for assessing urban tree condition through residuals derived from fitted growth models (Figure 6). Finally, there exists potential for further data integration. Here, we demonstrate a method for accurately detecting individual trees and associating them to tree entries in the city’s GIS database. Once the exact tree location has been determined using LiDAR, coregistered imagery can be used to
18
extract spectral information which, in combination with structural information gathered from LiDAR, can be used to produce composite indices of overall tree condition. 5.4 Use of a raster-based canopy height model As the availability of high-density LiDAR datasets rises and processing techniques develop, an increasing number of studies have attempted to perform both treetop detection and crown delineation directly on LiDAR point clouds, as opposed to rasterized surfaces such as a CHM (Li, Guo, Jakubowski, & Kelly, 2012; C. Zhang & Qiu, 2012). A common motive for this approach are the spatial errors that can be introduced when interpolating a point cloud to a grid (Smith, Holland, & Longley, 2004). Here, complex interpolation techniques were eschewed for a descriptive procedure by which CHM values were equal to the height above ground of the highest classified vegetation point within each cell. While the high point density of the dataset minimized concerns over artificial gaps in the CHM, this approach may still affect the results of the analysis. For instance, the spatial coordinates of the tree apexes are rounded to the nearest cell center, and the extent of some tree crowns may appear exaggerated. Tree height estimates derived from the CHM, however, will be identical to those taken from the raw LiDAR cloud, while density estimates, themselves calculated from the directly from the point cloud, are unaffected. 5.5 Limitations We recognize some limitations exist to this study. Our analysis is restricted to free-growing trees. Trees in dense clusters present challenges for treetop detection and crown delineation, while competition between trees can affect tree growth independently of environmental stressors. Furthermore, LiDAR is unable to measure a tree’s spectral properties, and so cannot detect signs of early senescence such as leaf yellowing. As such, LiDAR may not be an appropriate mechanism for detecting short term stressors such as insect outbreaks. The current high acquisition cost of LiDAR also makes it an impractical tool for continuous monitoring. Finally, ancillary data such as tree age and species are required for tree height and crown density to be meaningful indicators of condition. The lack of availability of this data would significantly reduce the ability of LiDAR to produce information on tree condition. 19
6. Conclusion This study demonstrates that high density, discrete return airborne LiDAR is a promising tool for evaluating the condition of individual free-growing urban trees. Using data from the GIS such as tree age and species allowed for tree-specific parameters to be used for an automated treetop detection algorithm, which was then applied to a sample of free-growing western redcedars. Establishing a connection between both datasets opens analytical opportunities such as using the residuals of growth models to assess tree condition. Using point cloud subsets corresponding to individual trees, two LiDAR metrics were tested as predictors for tree crown density. The percentage of non-ground returns, while conceptually sound, did not predict crown density reliably. The coefficient of variation of return height was able to predict tree crown density for trees over 8 m. Improved methods for estimating juvenile tree crown density are needed. As the availability of LiDAR data in urban areas increases, the demand to maximize its utility will rise. This study opens new applications for LiDAR in urban forestry by demonstrating its capacity to pinpoint exact tree locations, extract tree condition indicators, and augment existing urban tree datasets. This information can be used to produce comprehensive urban tree condition assessments which will, in turn, allow city managers to efficiently organize resources devoted to maintenance and investigate local factors impeding tree growth.
Acknowledgements Funding for this project was provided by an NSERC Discovery grant to Dr. Coops and Engage and Engage+ collaboration grants with Surrey City Energy. Andrew Plowright is supported by a NSERC CGS-M. We thank Mitchell Vartanian for help manually delineating trees and Curtis Chance for performing field work. The authors are grateful to Piotr Tompalski, Douglas Bolton and Gregory Rickbeil for advice on statistical analysis and data processing.
20
References BC Ministry of Forests. (2014). Site Tools 4.0. North Vancouver, BC: BC Ministry of Forests, Lands and Natural Resource Operations. Retrieved from https://www.for.gov.bc.ca/hts/growth/download/download.html Beucher, S., & Lantejoul, C. (1979). Use of watersheds in contour detection. In International Workshop on Image Processing: Real-time Edge and Motion Detection/Estimation. Rennes, France. Brandtberg, T., Warner, T. A., Landenberger, R. E., & McGraw, J. B. (2003). Detection and analysis of individual leaf-off tree crowns in small footprint, high sampling density lidar data from the eastern deciduous forest in North America. Remote Sensing of Environment, 85(3), 290–303. doi:10.1016/S0034-4257(03)00008-7 Chen, Q., Baldocchi, D., Gong, P., & Kelly, M. (2006). Isolating Individual Trees in a Savanna Woodland Using Small Footprint Lidar Data. Photogrammetric Engineering & Remote Sensing, 72(8), 923–932. doi:10.14358/PERS.72.8.923 City of Toronto. (2008). Ahead of the Storm: Preparing Toronto for Climate Change. Toronto, Ontario. Clutter, J. L., Fortson, J. C., Pienaar, L. V., Brister, G. H., & Bailey, R. L. (1983). Timber Management: A Quantitative Approach. New York, NY: John Wiley & Sons. Cohen, W. B., & Spies, T. a. (1992). Estimating structural attributes of Douglas-fir/western hemlock forest stands from Landsat and SPOT imagery. Remote Sensing of Environment, 41(1), 1–17. doi:10.1016/0034-4257(92)90056-P Coops, N. C., Hilker, T., Wulder, M. A., St-Onge, B., Newnham, G. J., Siggins, A., & Trofymow, J. A. (2007). Estimating canopy structure of Douglas-fir forest stands from discrete-return LiDAR. TreesStructure and Function, 21(3), 295–310. doi:10.1007/s00468-006-0119-6 Dobbertin, M. (2005). Tree growth as indicator of tree vitality and of tree reaction to environmental stress: A review. European Journal of Forest Research, 124(4), 319–333. doi:10.1007/s10342-0050085-3 Elzhov, T. V., Mullen, K. M., Spiess, A.-N., & Bolker, B. (2015). R interface to the Levenberg-Marquardt nonlinear least-squares algorithm found in MINPACK, plus support for bounds. R Package Version 1.1-8. Retrieved from http://cran.r-project.org/package=minpack.lm Fekedulegn, D., Siurtain, M. P. Mac, & Colbert, J. J. (1999). Parameter Estimation of Nonlinear Growth Models in Forestry. Silva Fennica, 33(4), 327–336. Goodwin, N., Coops, N. C., & Stone, C. (2005). Assessing plantation canopy condition from airborne imagery using spectral mixture analysis and fractional abundances. International Journal of Applied Earth Observation and Geoinformation, 7(1), 11–28. Retrieved from http://linkinghub.elsevier.com/retrieve/pii/S0303243404000832 Hyyppä, J., & Inkinen, M. (1999). Detecting and estimating attributes for single trees using laser scanner. The Photogrammetric Journal of Finland, 16(2). Jakubowski, M., Li, W., Guo, Q., & Kelly, M. (2013). Delineating Individual Trees from Lidar Data: A Comparison of Vector- and Raster-based Segmentation Approaches. Remote Sensing, 5(9), 4163– 4186. doi:10.3390/rs5094163 Jim, C. Y. (1993). Soil compaction as a constraint to tree growth in tropical & subtropical urban habitats. Environmental Conservation, 20(1), 35–49. 21
Jim, C. Y. (1998). Physical and chemical properties of a Hong Kong roadside soil in relation to urban tree growth. Urban Ecosystems, 2, 171–181. Kaartinen, H., Hyyppä, J., Yu, X., Vastaranta, M., Hyyppä, H., Kukko, A., … Wu, J.-C. (2012). An International Comparison of Individual Tree Detection and Extraction Using Airborne Laser Scanning. Remote Sensing, 4(12), 950–974. doi:10.3390/rs4040950 Kurucz, J. F. (1985). Metric SI tables for red cedar stands. Nanaimo, BC. Leckie, D. G., Gougeon, F., Hill, D., Quinn, R., Armstrong, L., & Shreenan, R. (2003). Combined highdensity lidar and multispectral imagery for individual tree crown analysis. Canadian Journal of Remote Sensing, 29(5), 633–649. Li, W., Guo, Q., Jakubowski, M. K., & Kelly, M. (2012). A New Method for Segmenting Individual Trees from the Lidar Point Cloud. Photogrammetric Engineering & Remote Sensing, 78(1), 75–84. doi:10.14358/PERS.78.1.75 Locke, D. H., Grove, J. M., Lu, J. W. T., Troy, A., Neil-dunne, J. P. M. O., & Beck, B. D. (2010). Prioritizing Preferable Locations for Increasing Urban Tree Canopy in New York City, 3(1), 1–18. Lovell, J., Jupp, D. L., Culvenor, D. S., & Coops, N. C. (2003). Using airborne and ground-based ranging lidar to measure canopy structure in Australian forests. Canadian Journal of Remote Sensing, 29(5), 607–622. Retrieved from http://scholar.google.com/scholar?hl=en&btnG=Search&q=intitle:Using+airborne+and+groundbased+ranging+lidar+to+measure+canopy+structure+in+Australian+forests#0 MacFaden, S. W., O’Neil-Dunne, J. P. M., Royar, A. R., Lu, J. W. T., & Rundle, A. G. (2012). High-resolution tree canopy mapping for New York City using LIDAR and object-based image analysis. Journal of Applied Remote Sensing, 6(1), 063567. doi:10.1117/1.JRS.6.063567 Malthus, T. J., & Younger, C. J. (2000). Remotely sensing stress in street trees using high spatial resolution imagery. In Proceedings of the 2nd International Conference on Geospatial Information in Agriculture and Forestry. Lake Buena Vista, Florida. McGaughey, R. J. (2014). FUSION/LDV: Software for LIDAR data analysis and visualization. United States Department of Agriculture, Forest Service. McPherson, E. G. (1993). Monitoring urban forest health. Environmental Monitoring and Assessment, 26(2-3), 165–174. McPherson, E. G., Simpson, J. R., Xiao, Q., & Wu, C. (2011). Million trees Los Angeles canopy cover and benefit assessment. Landscape and Urban Planning, 99(1), 40–50. doi:10.1016/j.landurbplan.2010.08.011 Næsset, E. (2002). Predicting forest stand characteristics with airborne scanning laser using a practical two-stage procedure and field data. Remote Sensing of Environment, 80(1), 88–99. doi:10.1016/S0034-4257(01)00290-5 Nowak, D. J., & McBride, J. R. (1991). Testing Microdensitometric Ability to Determine Monterey Pine Urban Tree. Photogrammetric Engineering & Remote Sensing, 59(1), 89–91. O’Neil-Dunne, J. P. M., MacFaden, S. W., Royar, A. R., & Pelletier, K. C. (2013). An object-based system for LiDAR data fusion and feature extraction. Geocarto International, 28(3), 227–242. doi:10.1080/10106049.2012.689015 Popescu, S. C., & Wynne, R. H. (2004). Seeing the trees in the forest: using lidar and multispectral data fusion with local filtering and variable window size for estimating tree height. Photogrammetric 22
Engineering and Remote, 70(5), 589–604. Retrieved from http://asprs.org/a/publications/pers/2004journal/may/2004_may_589-604.pdf Popescu, S. C., Wynne, R. H., & Nelson, R. H. (2002). Estimating plot-level tree heights with LIDAR: Local filtering with a canopy-height based variable window size. Computers and Electronics in Agriculture, 37(1-3), 71−95. Reich, R. W., & Price, R. (1999). Detection and classification of forest damage caused by tomentosus root rot using an airborne multispectral imager (CASI). In D. A. Hill & D. G. Leckie (Eds.), International Form on Automated Interpretation of High Spatial Resolution Digital Imagery for Forestry (pp. 179– 185). Victoria, BC. Sampson, P. H., Zarco-Tejada, P. J., Mohammed, G. H., Miller, J. R., & Noland, T. L. (2003). Hyperspectral remote sensing of forest condition: Estimating chlorophyll content in tolerant hardwoods. Forest Science, 49(3), 381–391. Schardt, M., Ziegler, M., Wimmer, A., & Wack, R. (2002). Assessment of forest parameters by means of laser scanning. International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences, 34(3/A), 302–309. Schomaker, M. E., Zarnoch, S. J., Bechtold, W. A., Latelle, D. J., Burkman, W. G., & Cox, S. M. (2007). Crown-Condition Classification : A Guide to Data Collection and Analysis. Schreyer, J., Tigges, J., Lakes, T., & Churkina, G. (2014). Using Airborne LiDAR and QuickBird Data for Modelling Urban Tree Carbon Storage and Its Distribution—A Case Study of Berlin. Remote Sensing, 6(11), 10636–10655. doi:10.3390/rs61110636 Singh, K. K., Davis, A. J., & Meentemeyer, R. K. (2015). Detecting understory plant invasion in urban forests using LiDAR. International Journal of Applied Earth Observation and Geoinformation, 38, 267–279. doi:10.1016/j.jag.2015.01.012 Smith, S., Holland, D., & Longley, P. (2004). The importance of understanding error in LiDAR digital elevation models. In 20th ISPRS conference. Istanbul, Turkey (pp. 12–34). Istanbul, Turkey. Retrieved from http://www.cartesia.org/geodoc/isprs2004/comm4/papers/488.pdf Statistics Canada. (2011). Surrey, British Columbia (Code 5915004) and British Columbia (Code 59) (table). Census Profile. 2011 Census. Ottawa. doi:98-316-XWE Stewart, H. (1984). Cedar: Tree of Life to the Northwest Coast Indians. Vancouver, British Columbia: Douglas & McIntyre. Stone, C., Coops, N., & Culvenor, D. (2000). Conceptual Development of a Eucalypt Canopy Condition Index Using High Resolution Spatial and Spectral Remote Sensing Imagery. Journal of Sustainable Forestry, 11(4), 23–45. doi:10.1300/J091v11n04_02 Suárez, J. C., Ontiveros, C., Smith, S., & Snape, S. (2005). The Use of Airborne LiDAR and Aerial Photography in the Estimation of Individual Tree Heights in Forestry. Computers & Geosciences, 31(2), 253–262. Tooke, T. R., Coops, N. C., & Voogt, J. (2009). Assessment of urban tree shade using fused LIDAR and high spatial resolution imagery. In Urban Remote Sensing Event, 2009 Joint (pp. 1–6). IEEE. Retrieved from http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5137619 Zarnoch, S. J., Bechtold, W. A., & Stolte, K. W. (2004). Using crown condition variables as indicators of forest health. Canadian Journal of Forest Research, 34(5), 1057–1070. doi:10.1139/X03-277 Zhang, C., & Qiu, F. (2012). Mapping individual tree species in an urban forest using airborne lidar data 23
and hyperspectral imagery. Photogrammetric Engineering & Remote Sensing, 78(10), 1079–1087. doi:10.14358/PERS.78.10.1079 Zhang, C., Zhou, Y., & Qiu, F. (2015). Individual Tree Segmentation from LiDAR Point Clouds for Urban Forest Inventory. Remote Sensing, 7(6), 7892–7913. doi:10.3390/rs70607892 Zhang, J., Sohn, G., & Brédif, M. (2014). A hybrid framework for single tree detection from airborne laser scanning data: A case study in temperate mature coniferous forests in Ontario, Canada. ISPRS Journal of Photogrammetry and Remote Sensing, 98, 44–57. doi:10.1016/j.isprsjprs.2014.08.007
24
Figure 1. Location of trees measured in the field within the city of Surrey, BC, Canada.
25
Figure 2. Process for locating the treetop of an entry in the city’s urban tree database. The white dotted line represents the approximate coordinates of the tree as registered within the GIS. A VWF is used to locate all potential treetops within a 20 m2 area centered on these coordinates. Using growth curve formulas, upper and lower height thresholds are computed according to the tree’s age. Finally, the closest treetop within these thresholds is selected.
Figure 3. Watershed segmentation: A) a CHM of two neighboring trees; B) the CHM is inverted so both trees resemble basins; C) the basins are filled with water from the bottom up; D) borders between the two tree crowns are drawn where the basins connect.
26
Figure 4. Rationale for using a resized crown outline when estimating crown density with the coefficient of variation of LiDAR point heights. Differentiating the vertical distribution of points between a sparse and a dense crown may be difficult when using a subset that extends to the edges of the tree (gray area). When the subset is limited to the tree’s apex and the interior of its crown (crosshatched area), a large amount of the returns being reflected from the tree’s exterior are excluded. Most of the remaining returns from a dense tree will occur near its apex, while a sparse tree, having allowed more returns to reach the interior of its crown, will have a more even distribution of return heights along its vertical profile.
27
Figure 5. Relationship between tree heights measured in the field and tree heights measured from the LiDAR data. Four outliers whose difference between field and LiDAR height was SD > 3 from the mean were removed.
28
Figure 6. Chapman-Richards height model fitted to the field trees. Many trees share the same planting year, resulting in clustering along the x-axis.
29
Figure 7. Examples of automatically generated (dotted lines) and manually delineated (solid lines) tree crown outlines. A) An accurate watershed segmentation outline (A = 0.92) B) A watershed segmentation error of commission: a neighboring tree or shrub has been included in the outline (A = 0.42). C) An accurate circular outline (A = 0.88). D) An inaccurate circular outline caused by an off-center tree apex (A = 0.39).
30
Figure 8. Ordinary least-squares linear regression models of field-measured crown density versus LiDAR metrics across multiple age classes. Left column: percentage of non-ground LiDAR returns. Right column: coefficient of variation of LiDAR heights.
31
Figure 9. A comparison of tree condition indicators for 40 field trees around Kildare Drive, in Surrey, BC. Gray points represent trees that are not western redcedars. A) Residuals against a model of tree growth for western redcedar. B) Estimated crown density.
32
Figure 10. Off-nadir LiDAR pulses may have unobstructed paths to the ground below an isolated tree in an open area (left). These pulses can be intercepted by neighbouring objects, reducing the number of returns to reach the ground below a tree in clustered surroundings (right).
33
Table 1. Number of reference trees per height class.
Height class interval [2 m, 5m) [5 m, 8 m) [8 m, 14 m)
Number of trees 34 76 59
34
S1618-8667(15)30064-9 http://dx.doi.org/doi:10.1016/j.ufug.2016.06.026 UFUG 25736
To appear in: Received date: Revised date: Accepted date:
7-10-2015 12-5-2016 30-6-2016
Please cite this article as: Plowright, Andrew A., Coops, Nicholas C., Eskelson, Bianca N.I., Sheppard, Stephen R.J., Aven, Neal W., Assessing urban tree condition using airborne light detection and ranging.Urban Forestry and Urban Greening http://dx.doi.org/10.1016/j.ufug.2016.06.026 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Assessing urban tree condition using airborne light detection and ranging Andrew A. Plowrighta M.Sc. candidate [email protected] * Corresponding author
Nicholas C. Coopsa Professor [email protected]
Bianca N.I. Eskelsona Assistant professor [email protected]
Stephen R.J. Shepparda Professor [email protected]
Neal W. Avenb Urban Forestry & Environmental Programs Manager [email protected]
a
University of British Columbia
Department of Forest Resources Management Faculty of Forestry Vancouver, Canada
b
City of Surrey
1
Urban Forestry and Environmental Programs Parks Division Surrey, Canada
Permanent address: IRSS: Integrated Remote Sensing Studio Department of Forest Resources Management Room 2231 2424 Main Mall Vancouver, BC, Canada V6T 1Z4
Highlights
Airborne LiDAR was used to estimate two indicators of urban tree condition: tree height and crown density. An automated method for detecting urban trees supported by GIS data was developed and tested. The coefficient of variation of LiDAR return heights is a stronger predictor of crown density than the percentage of non-ground returns. Improved methods for estimating crown density for juvenile trees are needed.
Abstract With increased interest in urban forests on behalf of city dwellers and urban planners, there is a growing need for comprehensive information on urban tree condition. This study examines the potential of airborne light detection and ranging (LiDAR) for evaluating tree condition in the urban center of Surrey, Canada. An approach to detecting and outlining free-growing trees from LiDAR data augmented by a municipal tree inventory was developed and validated. Once the trees were located, LiDAR was used to 2
estimate two field-measured indicators of tree condition: crown density and tree height. Tree heights estimated by LiDAR were, as expected, well correlated with field measurements (Pearson’s r = 0.927, p < 0.001), indicating accurate height estimates of successfully detected trees. Two LiDAR metrics, the percentage of non-ground LiDAR returns and the coefficient of variation of return height, were examined as predictors of crown density. While the percentage of non-ground returns performed relatively poorly (r2 between 0.005 and 0.23 across multiple tree height classes), the coefficient of variation of return height was able to predict crown density with an r2 = 0.617 for trees over 8 m. In addition, residuals derived using expected height growth from the known planting date and their LiDARderived height was found to be a useful tree condition metric. We conclude that despite the complexity of urban tree condition assessment, airborne LiDAR is a promising tool for detecting trees in an urban environment and measuring indicators of their condition.
Key words: LiDAR, GIS, remote sensing, urban trees, tree condition, tree detection
1. Introduction 1.1 The need for urban tree condition assessment Through their numerous aesthetic, environmental and economic benefits, trees are a key component to sustainable urban development. As the scientific foundation of urban forestry matures and the popular demand for green space increases, many city governments have undertaken ambitious projects to 3
expand their tree cover (City of Toronto, 2008; Locke et al., 2010; McPherson, Simpson, Xiao, & Wu, 2011). However, trees in urban environments are exposed to a wide variety of stress-inducing agents, such as unsuitable soil conditions, rerouted water flows, mechanical damage, and pollution (Jim, 1993, 1998; Malthus & Younger, 2000). As a result, urban forests are intensively managed, with expenditures towards urban trees per acre in the United States surpassing those in forested lands outside of cities (McPherson, 1993). With the considerable financial investment represented by urban forests, it is incumbent on city authorities to assess the health and vitality of the trees under their purview. Although hypothetical optimal tree vitality is difficult to define, the effects of environmental stressors are reflected by a range of clearly identifiable biochemical, physiological and structural symptoms (Dobbertin, 2005). The overall condition of a tree can be defined as the culmination of these symptoms (Stone, Coops, & Culvenor, 2000). Citywide data on tree condition can help optimize the allocation of resources and guide effective management prescriptions, such as targeted irrigation, pruning or removal, yet city authorities rarely have access to this type of information. Exhaustive ground surveys are prohibitively expensive, while sampling-based approaches may not be appropriate due to the wide range of tree species, age classes and the spatial heterogeneity of the urban environment (McPherson, 1993). 1.2 The use of remote sensing for assessing tree condition By providing efficient and repeatable means of acquiring quantitative and spatially explicit data, remote sensing may have the potential to address the need for comprehensive urban tree condition information. Given their capacity to detect changes in the reflectance of tree foliage, optical remote sensors have received considerable attention for this purpose (Goodwin, Coops, & Stone, 2005). Manifestations of tree stress, such as leaf chlorosis or necrosis, affect leaf spectral reflectance. Airborne multispectral imagery has been used to estimate chlorophyll content as an indicator of forest health
4
(Sampson, Zarco-Tejada, Mohammed, Miller, & Noland, 2003), detect root disease (Reich & Price, 1999), and assess tree condition in urban areas (Malthus & Younger, 2000; Nowak & McBride, 1991). Optical sensors are limited, however, when attempting to characterize a tree’s vertical structure. The dimensions and the vertical architecture of a tree are reflective of its net primary productivity and as such, are key indicators of its overall health and vigor (Schomaker et al., 2007). A large, dense crown is an indicator of optimal tree growth, while sparsely foliated crowns are a sign of deterioration and stress (Zarnoch, Bechtold, & Stolte, 2004). Tree height can also be used to assess a tree’s condition, as healthy trees grow rapidly while stressed trees experience stunted growth (Dobbertin, 2005). Although some vertical tree metrics can be estimated through indirect relationships with combinations of spectral bands (Cohen & Spies, 1992), airborne light detection and ranging (LiDAR) may provide the means to measure them directly. A LiDAR instrument emits pulses of light that are reflected off trees, ground surfaces, and other terrestrial features. Of particular note is LiDAR’s capacity to penetrate through gaps in the foliage, enabling it to directly measure the vertical aspects of tree crowns and forest canopies (Coops et al., 2007) While extensive research has demonstrated the utility of LiDAR in natural resource disciplines such as forestry, ecology, wildlife management and hydrology, its uses in urban forestry remain nascent. Zhang, Zhou, & Qiu (2015) developed an automated algorithm for detecting, outlining and mensurating urban trees directly from a LiDAR point cloud. Several studies have focused on integrating LiDAR with spectral imagery to map urban tree canopy cover (MacFaden, O’Neil-Dunne, Royar, Lu, & Rundle, 2012; O’NeilDunne, MacFaden, Royar, & Pelletier, 2013). LiDAR data can be particularly useful in dense urban areas where the shadowing effect of buildings make the use of spectral imagery problematic (MacFaden et al., 2012). Other applications of LiDAR and high resolution imagery include mapping urban tree species (C. Zhang & Qiu, 2012), modeling solar radiation effects (Tooke, Coops, & Voogt, 2009), estimating citywide
5
carbon storage (Schreyer, Tigges, Lakes, & Churkina, 2014) and detecting invasive plant species (Singh, Davis, & Meentemeyer, 2015). 1.3 Research objectives This paper explores the potential of airborne LiDAR as a tool for assessing urban tree condition within the city of Surrey, Canada. The objectives of this study are to: 1. Develop and evaluate an automated method for locating and outlining urban trees that is supported by ancillary GIS data collected by the city. 2. Examine LiDAR’s capacity to estimate two dendrological metrics: tree height and crown density. Methods for using these metrics as indicators of urban tree condition are presented and discussed. Our analysis focuses on western redcedar trees (Thuja plicata) located on city property. Western redcedar is the official tree of the Canadian province of British Columbia, and possesses unique cultural significance in the region. Furthermore, the species is of particular concern to city managers due to its preference for moist soils, and consequently its susceptibility to poor watering conditions (Stewart, 1984).
2. Data & study site 2.1 Study site The city of Surrey is located in the Greater Vancouver regional district, in the province of British Columbia, Canada. It is one of the fastest growing cities in Canada, with an 18.6% increase in population between 2006 and 2011 (Statistics Canada, 2011). Over 90,000 trees are actively managed on city property, with 3,500 to 5,000 additional trees being planted every year. As part of its annual tree maintenance budget, the city spends roughly 600,000 USD on watering alone. The current methods 6
employed for monitoring Surrey’s trees are mainly field-based, such as performing soil moisture spot checks in the summer drought months. 2.2 GIS database The city of Surrey maintains a comprehensive GIS database of all the trees that it plants and manages. Each entry includes a tree’s species, subspecies, planting date and approximate geographic coordinates. Tree coordinates are either recorded by field crews using GPS receivers, or, for trees predating the database’s creation, located through aerial photo interpretation. Due to inaccuracies inherent to mobile GPS units and the variable methods used for recording coordinates, certain entries contain positional errors, while others may be out-of-date or supply locations for trees that have been removed. 2.3 LiDAR data and derived products Airborne LiDAR data was acquired over Surrey by Airborne Imaging (Calgary, Alberta), under contract with the city in April, 2013. Trees were under leaf-off conditions. A Leica ALS70-HP discrete return LiDAR system, with up to four discrete returns per pulse, was flown at 1000 m above ground level with 688 m swaths with 50% overlap. The pulse rate was 500 KHz, which resulted in an average first-return density of 25 points per square meter. Before being delivered by the contractor, the raw LiDAR point cloud was classified into land class covers, such as ground, building or vegetation, using TerraScan software (TerraSolid Ltd., Helsinki). A 1 m2 rasterized digital elevation model (DEM) interpolated from classified ground points using a triangular irregular network (TIN) was also supplied.
7
3. Methods 3.1 Field data A ground survey was undertaken of 169 western redcedar trees, all of which had corresponding entries in the city’s database. The focus of this study was on detecting trees whose growth is affected by environmental stressors as opposed to competition with neighboring trees, and so the sample was restricted to free-growing trees. The criteria for defining a free-growing tree was that its crown not be in contact with the crown of any trees in its surroundings. Trees planted in parks, parking lots and along roadsides were included. In March, 2015, an initial four sampling sites were visited. For logistical reasons, only sites with high numbers of western redcedar were considered. Accompanying a surge in urban development, a substantial portion of the city’s western redcedar were planted in the late 1990s and early 2000s, and consequently, the four initial sites contained mostly trees planted within this period. To ensure that the sample represented a sufficiently broad range of ages, three additional sampling sites were subsequently visited, containing trees under 15 years and over 25 years. The combined sample from the seven sites contained 169 free-growing western redcedars with ages ranging from 11 to 33 years old (Figure 1). Surrey’s urban trees are approximately seven years old when planted. Tree heights were measured using a LTI TruPulse 360 laser rangefinder. Crown density was measured using methods described in Schomaker et al. (2007). Crown densities ranged from 25% to 95%, while heights ranged from 2.1 m to 14.5 m. 3.2 Canopy Height Model A canopy height model (CHM) was produced using FUSION software (McGaughey, 2014). This product is derived from the raw LiDAR data by recording the difference in elevation between the highest classified
8
vegetation return and the underlying DEM within each cell of a continuous grid. The maximum aboveground height of vegetation is indicated by the pixel value of the CHM at any given location. The selection of an appropriate cell size is an important parameter: large cells reduce variation within the CHM, while small cells can create gaps within the canopy and increase the volume of data. Chen, Baldocchi, Gong, & Kelly (2006) recommend calculating an optimal cell size based on the first-return density of the LiDAR data (Equation 1), where 𝜆 is returns/m2, and 𝑑 is cell size.
d
1
(1)
Using a density of 25 returns/m2, this equation yielded a cell size of 0.2 m. In practice, however, return density is not uniform, and so a conservative cell size of 0.5 m was selected to account for areas where flight swaths did not overlap. By computing the return count in areas free of water, it was found that approximately 99.94% of cells contained at least one return when using this resolution. 3.3 Locating trees For information on tree condition to be extracted from the LiDAR data, the location and spatial extent of the trees of interest are required. A substantial amount of research has been conducted on methods of locating and outlining individual trees from LiDAR data. Automated tree detection algorithms can either be applied directly to a point cloud or to a rasterized canopy height model (CHM) derived from raw LiDAR data (Jakubowski, Li, Guo, & Kelly, 2013). Approaches such as valley following and object-oriented classification have been applied to coregistered LiDAR data and multispectral imagery to detect and isolate individual trees in coniferous forests (Leckie et al., 2003; Suárez, Ontiveros, Smith, & Snape, 2005). Kaartinen et al. (2012) compared a variety of approaches of automated tree detection, and found that several outperformed manual processing of LiDAR data in terms of accuracy.
9
This study employed the popular and well-researched variable window filter (VWF) method for detecting trees (Popescu, Wynne, & Nelson, 2002; Popescu & Wynne, 2004). A VWF uses a moving focal window to scan the canopy for local canopy height maxima, which often correspond to treetops. When applied to a CHM, cells are tagged as treetops if their values are the highest within the window. Errors of commission occur when the window size is too small, causing branches, canopy protrusions, or other non-tree local maxima to be incorrectly tagged as treetops. Conversely, an overly large window may pass over treetops that are closely clustered together, causing errors of omission. A VWF minimizes both types of error by adjusting the size of the window according to the height of the cell on which it is centered. This is based on the relationship between tree height and crown width: pixels with high values are likely to belong to tall trees with correspondingly wide crowns, and vice versa. To parameterize the VWF, a formula relating tree height and crown width is required. Crown outlines for a randomly selected subset of 115 surveyed trees were manually delineated through visual interpretation of the CHM. Using a quadratic model to regress the average crown diameter of these reference outlines with their height produced Equation 2, which yielded an adjusted r2 value of 0.5958. ̂𝑊𝑖𝑛𝑑𝑜𝑤 when centered over The equation is then used to derive the diameter of the search window 𝐷 any given CHM cell with a height value of HCell. 2 Dˆ W in d o w 0 .6 9 9 0 6 0 .6 6 5 1 1 H C e ll 0 .0 1 0 5 2 H C e ll
(2)
While Surrey’s GIS database contains the approximate locations of all the trees maintained by the city, coordinates may be offset from the trees’ actual locations, and some trees may be missing entirely. The following process (Figure 2) uses a VWF to complement the city’s GIS data and locate trees with increased precision. A 20 m2 square search area, approximately the size of the largest possible tree crown for western redcedar, is centered on a tree’s coordinates as recorded in the database. To prevent the detection of 10
false treetops from shrubs or other low-lying non-tree objects, all CHM pixels with a height value less than 1 m are masked out within the search area. Then, the VWF is applied to locate all potential treetops within the tree’s immediate vicinity. The potential treetops identified by the VWF are then filtered according to height. Commonly used growth curves for coastal western redcedar developed by Kurucz (1985) provide estimates of a tree’s height over time for any given site index. Site indices, which are expressed as the height of a dominant tree in meters at age 50, measure the growing potential of a forested site, with an elevated site index indicating optimal growing conditions and vice versa (Clutter, Fortson, Pienaar, Brister, & Bailey, 1983). Site indices of 40 and 20 were selected as upper and lower thresholds, which correspond to the range found within the field sample and are considered to represent the range of growing conditions found in the city. Using Site Tools software (BC Ministry of Forests, 2014), maximum and minimum values HTreetopMax and HTreetopMin were computed for each tree according to its age. Treetops outside of these thresholds were removed, and the remaining treetop closest to the tree’s coordinates was then selected. If no treetops within these thresholds are found, the tree is reported as missing. 3.4 Outlining tree crowns Once the coordinates of a given tree have been located, an outline of its crown’s horizontal extent is needed. An image processing technique known as watershed segmentation, originally developed for delineating drainage basins (Beucher & Lantejoul, 1979), can be used to segment tree crowns due to the morphological similarity between canopies and topographical terrain models (Figure 3). To avoid oversegmentation, wherein branches or other protrusions in the canopy are assigned spurious segments, the process is constrained by the pre-defined locations of treetops as detected by the VWF (Chen et al., 2006; Schardt, Ziegler, Wimmer, & Wack, 2002; J. Zhang, Sohn, & Brédif, 2014).
11
A second automated outlining approach, wherein the crown’s horizontal size and shape are approximated using a circle, was also tested. The average crown width of each tree is computed in a process described by McGaughey (2014). This approach determines the crown radius along 16 equally spaced radial profiles centered on a tree’s apex. Along each profile, the radius is given by the distance between the apex and the first CHM pixel that is either A) a local minima within the profile or B) less than 50% of the tree’s height. A circular outline is then generated using this average of these radii. Both automated outlining approaches were compared. Brandtberg et al. (2003) developed a fuzzy method for assessing the similarity between two corresponding tree crown outlines. This approach, which uses spatial weighting to assign greater importance to the center of an outline than its periphery, produces a segmentation assessment value (A) between two outlines, with an A = 1 indicating a perfect match, and any other case an A < 1. The ratio between the areas of automated outlines and their corresponding reference outlines (Q) was also computed. 3.5 Estimating crown density In a technique similar to the use of a cookie-cutter, LiDAR point cloud subsets corresponding to individual trees were created according to the boundaries of the crown outlines. The vertical measurements of the points in these subset were then normalized according to the DEM. Two metrics were tested as potential predictors of crown density. Previous attempts at estimating the foliar density of forests at an area-based level have used the ratio of LiDAR returns below a certain height threshold over the total number of returns (Lovell, Jupp, Culvenor, & Coops, 2003). The reasoning behind this metric is that denser canopies are more likely to intercept LiDAR pulses, preventing a larger portion of them from reaching the ground. Here, the percentage of returns with a height above ground higher than 0.5 m within each tree’s normalized LiDAR point cloud subset will be referred to as “percentage of non-ground returns”. 12
Other studies have used the coefficient of variation (CV) of the height of LiDAR returns to estimate forest attributes (Næsset, 2002). A large CV indicates a wide dispersion of LiDAR returns along the vertical profile, suggesting increased penetration of the canopy. While this may hold true for an area of continuous forest, certain considerations should be made for an individual tree. For instance, in an open area where the laser path is unobstructed by neighboring objects, the conical shape of the western redcedar may intercept LiDAR pulses evenly along the vertical profile, regardless of the crown’s density (Figure 4). To correct for this, the size of the outline used to subset the LiDAR points was halved, so that a large portion of the returns from the crown’s exterior are excluded. The coefficient of variation of the height of the returns within this subset will be referred to as “CV of return heights”. To remove the potential confounding effects of tree height, the reference trees were separated into 3 m equal interval height classes spanning the range of tree heights as measured in the field (Table 1). Due to the small number of trees in the highest class (14 trees between 11 m and 14 m) it was merged with the preceding class to ensure that all classes had a minimum of 30 trees. 3.6 Tree height as an indicator of condition The highest point within each LiDAR point subset was used as an estimate of the tree’s height. For a tree’s height to be an informative measure of its condition, it needs to be compared to a reference of growth for that tree’s species. With the availability of planting dates from Surrey’s GIS database, it is possible to fit growth models to trees whose heights have been accurately measured using LiDAR. The residual of a tree’s height against this model can then be used as an indicator of the tree’s condition (Dobbertin, 2005). Trees with positive residuals are growing at an optimum rate, while trees with negative residuals may be experiencing some sort of stress that is impeding growth. Based on methods described in Fekedulegn et al. (1999), a Chapman-Richards model was fit to the field-measured trees using the minpack.lm package for R statistical software (Elzhov, Mullen, Spiess, & Bolker, 2015). The
13
model’s general formula is given by Equation 3, where Hi is the height of a given tree (i), agei is its age according to the city’s database, εi is the model’s residual and 𝛽 are estimated coefficients:
H i 1 1 e
2 a g ei
3
i
(3)
4. Results 4.1 Treetop location and tree height estimation Of the 169 reference trees measured in the field, the automated tree detection algorithm reported one as missing. Figure 5 shows the relationship between tree heights measured in the field and heights measured from the LiDAR data, which had a Pearson correlation coefficient of r = 0.927, p < 0.001. Trees for which the difference between LiDAR- and field-based height measurements was more than three standard deviations from the mean were selected for visual inspection. The inspection revealed that these four outliers had been incorrectly identified by the algorithm and were removed. The remaining LiDAR heights were, on average, 0.72 m lower than those measured in the field, with a standard deviation of 1.09 m. Figure 6 shows the Chapman-Richards growth model fitted to the 169 field trees using their LiDAR-based height measurements. 4.2 Tree crown outlines On average, the ratio between the reference and the watershed segmentation outline areas (Q) was Q = 1.06, and Q = 1.01 for the circular outlines. Manually delineated reference outlines produced a segmentation assessment value (A) of A = 0.71 when compared to the watershed segmentation outlines and A = 0.63 when compared to the circular outlines. Figure 7 shows examples of various segmentation assessment values A. The standard deviation of A for watershed segmentation outlines was 0.21, with 19% of the outlines being nearly identical to their references (A > 0.9), and 23% of the outlines being
14
substantially different (A < 0.5). In comparison, the standard deviation for circular outlines was 0.14, with no outlines at A > 0.9, but only 17% with A < 0.5. 4.3 Crown density estimation Figure 8 shows ordinary least-squares linear regression models with field-measured crown density as the response variable and the percentage of non-ground LiDAR returns and the CV of return heights as separate predictor variables. For each of the three height classes, CV of return heights was a better predictor based on r2 values. Both predictor variables performed best in the highest height class (8 m to 14 m) and poorest in the smallest height class (2 m to 5 m). With an r2 = 0.617, predictions were best for the 8 m to 14 m height class using the CV of return heights. In order to demonstrate how the estimates of tree condition could be applied over the urban landscape, Figure 9 shows a comparison between tree height residuals and crown density for 40 of the field-measured trees in a subset of the residential area of Surrey. The map shows some apparent clumping in the spatial arrangement of tree condition indicators. Investigation of the drivers of these spatial patterns is the logical subsequent analysis which will follow in future work.
5. Discussion 5.1 Accuracy of treetop location algorithm For an automated algorithm to accurately estimate crown density from LiDAR data at an individual tree basis, the precise location of that tree is required. It should not be expected that city GIS databases, when available, will contain tree coordinates that meet this level of precision. Therefore, it is important to incorporate steps that will account for this potential inaccuracy, both by correcting positional errors and by skipping trees that may have been removed or incorrectly recorded.
15
Here, a VWF was applied to identify all potential treetops in the vicinity of a tree’s entry in the city’s GIS database. When used in forestry research, tree extraction routines such as the VWF are generally applied over large areas using a single set of parameters. These parameters are calibrated according to the forest’s species and age composition. This task is challenging in urban forests, however, whose highly variable structure make parameterizing automated algorithms difficult. Surrey’s existing GIS data can be used to address this issue. Although we focus here on a single species (Thuja plicata), speciesspecific equations for deriving search window sizes (such as Equation 2) can be developed and applied when extending this approach to other types of trees. Once a series of potential treetops have been identified, the treetop matching the tree entry must be selected. Using growth curves to compute a range of likely heights based on the tree’s age, and then using this range to filter potential treetops identified by the VWF was found to be a reliable way of locating trees within the LiDAR data and avoiding the incorrect selection of neighboring trees or false treetops such as buildings or power lines. The efficacy of this method was demonstrated by the highly significant correlation between height estimates obtained from the LiDAR data and heights measured in the field. 5.2 Performance of crown outline methods In comparison with previous methodologically similar studies (Brandtberg, Warner, Landenberger, & McGraw, 2003), both automated approaches for generating tree crown outlines produced acceptable results. The average ratio between the area of the circular outlines and their corresponding references (Q = 1.02) suggests this is an accurate method for measuring crown diameter. This approach produced a lower average segmentation assessment value (A = 0.63) than the watershed segmentation algorithm (A = 0.71), which is likely due to a mismatch between the circles and the sinuated contour of the tree’s fringe. It should be noted that western redcedars generally have a circular shape, and that the quality of
16
a circular outline will likely be reduced for trees with irregularly shaped crowns, and particularly those with off-center apexes. The standard deviation (SD) of A was highest for watershed segmentation outlines (SD = 0.21) which indicates a wider variation in the quality of the outlines. This is explained by this method’s reliance on accurate prior treetop detection. When all treetops are successfully detected within the vicinity of the targeted tree, watershed segmentation has the potential to produce highly accurate outlines, as demonstrated by the 19% of outlines with an A > 0.9. Conversely, watershed segmentation can oversegment a single tree when spurious treetops are mistakenly identified within its crown, or include multiple tree crowns in a single outline when treetops are missed. Although the watershed segmentation algorithm outperformed the circular outlines with respect to the average A, its higher incidence of unreliable outlines (23% of watershed segmentation outlines with A < 0.5, compared to 17% for circular outlines), has important implications when attempting to extract LiDAR metrics. For instance, an outline that erroneously includes neighboring trees will result in a highly inaccurate LiDAR-based estimate of crown density. Furthermore, Figure 4 illustrates the rationale for using only the horizontal center of a tree’s crown for density estimation. For this reason, the circular outlines were used to estimate crown density in lieu of those generated using watershed segmentation. 5.3 Potential use of LiDAR metrics for assessing urban tree condition While previous studies have used the ratio of LiDAR returns intercepted over a certain height (usually between 0.5 m to 2 m) to measure foliar density over large areas, this method does not appear to be applicable at the individual tree level. The percentage of returns above 0.5 m was a poor predictor of crown density for all tree height classes. Potential reasons include varying live crown ratios (the ratio of the vertical length of the crown to the tree’s full height) or interference from surrounding objects.
17
Figure 10 illustrates how neighboring trees or buildings could potentially affect the number of returns to reach the ground under a tree. The coefficient of variation of LiDAR return heights is a stronger predictor of density, particularly for trees in the 8 m to 14 m height class (r2 = 0.617). Taller trees intercept a greater number of LiDAR returns, which may reduce variability and improve the relationship between the two variables. According to estimates made by applying the treetop detection algorithm citywide, approximately 69% of Surrey’s urban western redcedars are taller than 8 m, suggesting that the coefficient of variation of return heights may be a useful measure of crown density for a large portion of Surrey’s urban trees. It should be noted, however, that the coefficient of variation is influenced by the number of LiDAR returns per tree and so if this metrics is applied in future work, transformations that account for variable LiDAR return density levels should be applied. Previous studies have established LiDAR as a reliable tool for measuring individual tree heights (Hyyppä & Inkinen, 1999). LiDAR measurements tend to underestimate tree height, however, as pulses have little chance of hitting a tree’s apex. Tree growth occurring between LiDAR and field measurements may have also contributed to this effect. These underestimations are generally consistent, as they are here by an average of 0.72 m, which allows them to potentially be corrected using empirically-derived offsets (Schardt et al., 2002). The correlation between tree heights estimated by LiDAR and heights measured in the field demonstrate both the functionality of the treetop detection method and the accuracy of LiDAR height estimates. By extension, these accurate height estimates support this approach’s potential for assessing urban tree condition through residuals derived from fitted growth models (Figure 6). Finally, there exists potential for further data integration. Here, we demonstrate a method for accurately detecting individual trees and associating them to tree entries in the city’s GIS database. Once the exact tree location has been determined using LiDAR, coregistered imagery can be used to
18
extract spectral information which, in combination with structural information gathered from LiDAR, can be used to produce composite indices of overall tree condition. 5.4 Use of a raster-based canopy height model As the availability of high-density LiDAR datasets rises and processing techniques develop, an increasing number of studies have attempted to perform both treetop detection and crown delineation directly on LiDAR point clouds, as opposed to rasterized surfaces such as a CHM (Li, Guo, Jakubowski, & Kelly, 2012; C. Zhang & Qiu, 2012). A common motive for this approach are the spatial errors that can be introduced when interpolating a point cloud to a grid (Smith, Holland, & Longley, 2004). Here, complex interpolation techniques were eschewed for a descriptive procedure by which CHM values were equal to the height above ground of the highest classified vegetation point within each cell. While the high point density of the dataset minimized concerns over artificial gaps in the CHM, this approach may still affect the results of the analysis. For instance, the spatial coordinates of the tree apexes are rounded to the nearest cell center, and the extent of some tree crowns may appear exaggerated. Tree height estimates derived from the CHM, however, will be identical to those taken from the raw LiDAR cloud, while density estimates, themselves calculated from the directly from the point cloud, are unaffected. 5.5 Limitations We recognize some limitations exist to this study. Our analysis is restricted to free-growing trees. Trees in dense clusters present challenges for treetop detection and crown delineation, while competition between trees can affect tree growth independently of environmental stressors. Furthermore, LiDAR is unable to measure a tree’s spectral properties, and so cannot detect signs of early senescence such as leaf yellowing. As such, LiDAR may not be an appropriate mechanism for detecting short term stressors such as insect outbreaks. The current high acquisition cost of LiDAR also makes it an impractical tool for continuous monitoring. Finally, ancillary data such as tree age and species are required for tree height and crown density to be meaningful indicators of condition. The lack of availability of this data would significantly reduce the ability of LiDAR to produce information on tree condition. 19
6. Conclusion This study demonstrates that high density, discrete return airborne LiDAR is a promising tool for evaluating the condition of individual free-growing urban trees. Using data from the GIS such as tree age and species allowed for tree-specific parameters to be used for an automated treetop detection algorithm, which was then applied to a sample of free-growing western redcedars. Establishing a connection between both datasets opens analytical opportunities such as using the residuals of growth models to assess tree condition. Using point cloud subsets corresponding to individual trees, two LiDAR metrics were tested as predictors for tree crown density. The percentage of non-ground returns, while conceptually sound, did not predict crown density reliably. The coefficient of variation of return height was able to predict tree crown density for trees over 8 m. Improved methods for estimating juvenile tree crown density are needed. As the availability of LiDAR data in urban areas increases, the demand to maximize its utility will rise. This study opens new applications for LiDAR in urban forestry by demonstrating its capacity to pinpoint exact tree locations, extract tree condition indicators, and augment existing urban tree datasets. This information can be used to produce comprehensive urban tree condition assessments which will, in turn, allow city managers to efficiently organize resources devoted to maintenance and investigate local factors impeding tree growth.
Acknowledgements Funding for this project was provided by an NSERC Discovery grant to Dr. Coops and Engage and Engage+ collaboration grants with Surrey City Energy. Andrew Plowright is supported by a NSERC CGS-M. We thank Mitchell Vartanian for help manually delineating trees and Curtis Chance for performing field work. The authors are grateful to Piotr Tompalski, Douglas Bolton and Gregory Rickbeil for advice on statistical analysis and data processing.
20
References BC Ministry of Forests. (2014). Site Tools 4.0. North Vancouver, BC: BC Ministry of Forests, Lands and Natural Resource Operations. Retrieved from https://www.for.gov.bc.ca/hts/growth/download/download.html Beucher, S., & Lantejoul, C. (1979). Use of watersheds in contour detection. In International Workshop on Image Processing: Real-time Edge and Motion Detection/Estimation. Rennes, France. Brandtberg, T., Warner, T. A., Landenberger, R. E., & McGraw, J. B. (2003). Detection and analysis of individual leaf-off tree crowns in small footprint, high sampling density lidar data from the eastern deciduous forest in North America. Remote Sensing of Environment, 85(3), 290–303. doi:10.1016/S0034-4257(03)00008-7 Chen, Q., Baldocchi, D., Gong, P., & Kelly, M. (2006). Isolating Individual Trees in a Savanna Woodland Using Small Footprint Lidar Data. Photogrammetric Engineering & Remote Sensing, 72(8), 923–932. doi:10.14358/PERS.72.8.923 City of Toronto. (2008). Ahead of the Storm: Preparing Toronto for Climate Change. Toronto, Ontario. Clutter, J. L., Fortson, J. C., Pienaar, L. V., Brister, G. H., & Bailey, R. L. (1983). Timber Management: A Quantitative Approach. New York, NY: John Wiley & Sons. Cohen, W. B., & Spies, T. a. (1992). Estimating structural attributes of Douglas-fir/western hemlock forest stands from Landsat and SPOT imagery. Remote Sensing of Environment, 41(1), 1–17. doi:10.1016/0034-4257(92)90056-P Coops, N. C., Hilker, T., Wulder, M. A., St-Onge, B., Newnham, G. J., Siggins, A., & Trofymow, J. A. (2007). Estimating canopy structure of Douglas-fir forest stands from discrete-return LiDAR. TreesStructure and Function, 21(3), 295–310. doi:10.1007/s00468-006-0119-6 Dobbertin, M. (2005). Tree growth as indicator of tree vitality and of tree reaction to environmental stress: A review. European Journal of Forest Research, 124(4), 319–333. doi:10.1007/s10342-0050085-3 Elzhov, T. V., Mullen, K. M., Spiess, A.-N., & Bolker, B. (2015). R interface to the Levenberg-Marquardt nonlinear least-squares algorithm found in MINPACK, plus support for bounds. R Package Version 1.1-8. Retrieved from http://cran.r-project.org/package=minpack.lm Fekedulegn, D., Siurtain, M. P. Mac, & Colbert, J. J. (1999). Parameter Estimation of Nonlinear Growth Models in Forestry. Silva Fennica, 33(4), 327–336. Goodwin, N., Coops, N. C., & Stone, C. (2005). Assessing plantation canopy condition from airborne imagery using spectral mixture analysis and fractional abundances. International Journal of Applied Earth Observation and Geoinformation, 7(1), 11–28. Retrieved from http://linkinghub.elsevier.com/retrieve/pii/S0303243404000832 Hyyppä, J., & Inkinen, M. (1999). Detecting and estimating attributes for single trees using laser scanner. The Photogrammetric Journal of Finland, 16(2). Jakubowski, M., Li, W., Guo, Q., & Kelly, M. (2013). Delineating Individual Trees from Lidar Data: A Comparison of Vector- and Raster-based Segmentation Approaches. Remote Sensing, 5(9), 4163– 4186. doi:10.3390/rs5094163 Jim, C. Y. (1993). Soil compaction as a constraint to tree growth in tropical & subtropical urban habitats. Environmental Conservation, 20(1), 35–49. 21
Jim, C. Y. (1998). Physical and chemical properties of a Hong Kong roadside soil in relation to urban tree growth. Urban Ecosystems, 2, 171–181. Kaartinen, H., Hyyppä, J., Yu, X., Vastaranta, M., Hyyppä, H., Kukko, A., … Wu, J.-C. (2012). An International Comparison of Individual Tree Detection and Extraction Using Airborne Laser Scanning. Remote Sensing, 4(12), 950–974. doi:10.3390/rs4040950 Kurucz, J. F. (1985). Metric SI tables for red cedar stands. Nanaimo, BC. Leckie, D. G., Gougeon, F., Hill, D., Quinn, R., Armstrong, L., & Shreenan, R. (2003). Combined highdensity lidar and multispectral imagery for individual tree crown analysis. Canadian Journal of Remote Sensing, 29(5), 633–649. Li, W., Guo, Q., Jakubowski, M. K., & Kelly, M. (2012). A New Method for Segmenting Individual Trees from the Lidar Point Cloud. Photogrammetric Engineering & Remote Sensing, 78(1), 75–84. doi:10.14358/PERS.78.1.75 Locke, D. H., Grove, J. M., Lu, J. W. T., Troy, A., Neil-dunne, J. P. M. O., & Beck, B. D. (2010). Prioritizing Preferable Locations for Increasing Urban Tree Canopy in New York City, 3(1), 1–18. Lovell, J., Jupp, D. L., Culvenor, D. S., & Coops, N. C. (2003). Using airborne and ground-based ranging lidar to measure canopy structure in Australian forests. Canadian Journal of Remote Sensing, 29(5), 607–622. Retrieved from http://scholar.google.com/scholar?hl=en&btnG=Search&q=intitle:Using+airborne+and+groundbased+ranging+lidar+to+measure+canopy+structure+in+Australian+forests#0 MacFaden, S. W., O’Neil-Dunne, J. P. M., Royar, A. R., Lu, J. W. T., & Rundle, A. G. (2012). High-resolution tree canopy mapping for New York City using LIDAR and object-based image analysis. Journal of Applied Remote Sensing, 6(1), 063567. doi:10.1117/1.JRS.6.063567 Malthus, T. J., & Younger, C. J. (2000). Remotely sensing stress in street trees using high spatial resolution imagery. In Proceedings of the 2nd International Conference on Geospatial Information in Agriculture and Forestry. Lake Buena Vista, Florida. McGaughey, R. J. (2014). FUSION/LDV: Software for LIDAR data analysis and visualization. United States Department of Agriculture, Forest Service. McPherson, E. G. (1993). Monitoring urban forest health. Environmental Monitoring and Assessment, 26(2-3), 165–174. McPherson, E. G., Simpson, J. R., Xiao, Q., & Wu, C. (2011). Million trees Los Angeles canopy cover and benefit assessment. Landscape and Urban Planning, 99(1), 40–50. doi:10.1016/j.landurbplan.2010.08.011 Næsset, E. (2002). Predicting forest stand characteristics with airborne scanning laser using a practical two-stage procedure and field data. Remote Sensing of Environment, 80(1), 88–99. doi:10.1016/S0034-4257(01)00290-5 Nowak, D. J., & McBride, J. R. (1991). Testing Microdensitometric Ability to Determine Monterey Pine Urban Tree. Photogrammetric Engineering & Remote Sensing, 59(1), 89–91. O’Neil-Dunne, J. P. M., MacFaden, S. W., Royar, A. R., & Pelletier, K. C. (2013). An object-based system for LiDAR data fusion and feature extraction. Geocarto International, 28(3), 227–242. doi:10.1080/10106049.2012.689015 Popescu, S. C., & Wynne, R. H. (2004). Seeing the trees in the forest: using lidar and multispectral data fusion with local filtering and variable window size for estimating tree height. Photogrammetric 22
Engineering and Remote, 70(5), 589–604. Retrieved from http://asprs.org/a/publications/pers/2004journal/may/2004_may_589-604.pdf Popescu, S. C., Wynne, R. H., & Nelson, R. H. (2002). Estimating plot-level tree heights with LIDAR: Local filtering with a canopy-height based variable window size. Computers and Electronics in Agriculture, 37(1-3), 71−95. Reich, R. W., & Price, R. (1999). Detection and classification of forest damage caused by tomentosus root rot using an airborne multispectral imager (CASI). In D. A. Hill & D. G. Leckie (Eds.), International Form on Automated Interpretation of High Spatial Resolution Digital Imagery for Forestry (pp. 179– 185). Victoria, BC. Sampson, P. H., Zarco-Tejada, P. J., Mohammed, G. H., Miller, J. R., & Noland, T. L. (2003). Hyperspectral remote sensing of forest condition: Estimating chlorophyll content in tolerant hardwoods. Forest Science, 49(3), 381–391. Schardt, M., Ziegler, M., Wimmer, A., & Wack, R. (2002). Assessment of forest parameters by means of laser scanning. International Archives of Photogrammetry Remote Sensing and Spatial Information Sciences, 34(3/A), 302–309. Schomaker, M. E., Zarnoch, S. J., Bechtold, W. A., Latelle, D. J., Burkman, W. G., & Cox, S. M. (2007). Crown-Condition Classification : A Guide to Data Collection and Analysis. Schreyer, J., Tigges, J., Lakes, T., & Churkina, G. (2014). Using Airborne LiDAR and QuickBird Data for Modelling Urban Tree Carbon Storage and Its Distribution—A Case Study of Berlin. Remote Sensing, 6(11), 10636–10655. doi:10.3390/rs61110636 Singh, K. K., Davis, A. J., & Meentemeyer, R. K. (2015). Detecting understory plant invasion in urban forests using LiDAR. International Journal of Applied Earth Observation and Geoinformation, 38, 267–279. doi:10.1016/j.jag.2015.01.012 Smith, S., Holland, D., & Longley, P. (2004). The importance of understanding error in LiDAR digital elevation models. In 20th ISPRS conference. Istanbul, Turkey (pp. 12–34). Istanbul, Turkey. Retrieved from http://www.cartesia.org/geodoc/isprs2004/comm4/papers/488.pdf Statistics Canada. (2011). Surrey, British Columbia (Code 5915004) and British Columbia (Code 59) (table). Census Profile. 2011 Census. Ottawa. doi:98-316-XWE Stewart, H. (1984). Cedar: Tree of Life to the Northwest Coast Indians. Vancouver, British Columbia: Douglas & McIntyre. Stone, C., Coops, N., & Culvenor, D. (2000). Conceptual Development of a Eucalypt Canopy Condition Index Using High Resolution Spatial and Spectral Remote Sensing Imagery. Journal of Sustainable Forestry, 11(4), 23–45. doi:10.1300/J091v11n04_02 Suárez, J. C., Ontiveros, C., Smith, S., & Snape, S. (2005). The Use of Airborne LiDAR and Aerial Photography in the Estimation of Individual Tree Heights in Forestry. Computers & Geosciences, 31(2), 253–262. Tooke, T. R., Coops, N. C., & Voogt, J. (2009). Assessment of urban tree shade using fused LIDAR and high spatial resolution imagery. In Urban Remote Sensing Event, 2009 Joint (pp. 1–6). IEEE. Retrieved from http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5137619 Zarnoch, S. J., Bechtold, W. A., & Stolte, K. W. (2004). Using crown condition variables as indicators of forest health. Canadian Journal of Forest Research, 34(5), 1057–1070. doi:10.1139/X03-277 Zhang, C., & Qiu, F. (2012). Mapping individual tree species in an urban forest using airborne lidar data 23
and hyperspectral imagery. Photogrammetric Engineering & Remote Sensing, 78(10), 1079–1087. doi:10.14358/PERS.78.10.1079 Zhang, C., Zhou, Y., & Qiu, F. (2015). Individual Tree Segmentation from LiDAR Point Clouds for Urban Forest Inventory. Remote Sensing, 7(6), 7892–7913. doi:10.3390/rs70607892 Zhang, J., Sohn, G., & Brédif, M. (2014). A hybrid framework for single tree detection from airborne laser scanning data: A case study in temperate mature coniferous forests in Ontario, Canada. ISPRS Journal of Photogrammetry and Remote Sensing, 98, 44–57. doi:10.1016/j.isprsjprs.2014.08.007
24
Figure 1. Location of trees measured in the field within the city of Surrey, BC, Canada.
25
Figure 2. Process for locating the treetop of an entry in the city’s urban tree database. The white dotted line represents the approximate coordinates of the tree as registered within the GIS. A VWF is used to locate all potential treetops within a 20 m2 area centered on these coordinates. Using growth curve formulas, upper and lower height thresholds are computed according to the tree’s age. Finally, the closest treetop within these thresholds is selected.
Figure 3. Watershed segmentation: A) a CHM of two neighboring trees; B) the CHM is inverted so both trees resemble basins; C) the basins are filled with water from the bottom up; D) borders between the two tree crowns are drawn where the basins connect.
26
Figure 4. Rationale for using a resized crown outline when estimating crown density with the coefficient of variation of LiDAR point heights. Differentiating the vertical distribution of points between a sparse and a dense crown may be difficult when using a subset that extends to the edges of the tree (gray area). When the subset is limited to the tree’s apex and the interior of its crown (crosshatched area), a large amount of the returns being reflected from the tree’s exterior are excluded. Most of the remaining returns from a dense tree will occur near its apex, while a sparse tree, having allowed more returns to reach the interior of its crown, will have a more even distribution of return heights along its vertical profile.
27
Figure 5. Relationship between tree heights measured in the field and tree heights measured from the LiDAR data. Four outliers whose difference between field and LiDAR height was SD > 3 from the mean were removed.
28
Figure 6. Chapman-Richards height model fitted to the field trees. Many trees share the same planting year, resulting in clustering along the x-axis.
29
Figure 7. Examples of automatically generated (dotted lines) and manually delineated (solid lines) tree crown outlines. A) An accurate watershed segmentation outline (A = 0.92) B) A watershed segmentation error of commission: a neighboring tree or shrub has been included in the outline (A = 0.42). C) An accurate circular outline (A = 0.88). D) An inaccurate circular outline caused by an off-center tree apex (A = 0.39).
30
Figure 8. Ordinary least-squares linear regression models of field-measured crown density versus LiDAR metrics across multiple age classes. Left column: percentage of non-ground LiDAR returns. Right column: coefficient of variation of LiDAR heights.
31
Figure 9. A comparison of tree condition indicators for 40 field trees around Kildare Drive, in Surrey, BC. Gray points represent trees that are not western redcedars. A) Residuals against a model of tree growth for western redcedar. B) Estimated crown density.
32
Figure 10. Off-nadir LiDAR pulses may have unobstructed paths to the ground below an isolated tree in an open area (left). These pulses can be intercepted by neighbouring objects, reducing the number of returns to reach the ground below a tree in clustered surroundings (right).
33
Table 1. Number of reference trees per height class.
Height class interval [2 m, 5m) [5 m, 8 m) [8 m, 14 m)
Number of trees 34 76 59
34