Urban-Tree-Attribute Update Using Multisource Single-Tree Inventory

The requirements for up-to-date tree data in city parks and forests are increasing, and an important question is how to keep the digital databases current for various applications. Traditional map-updating procedures, such as visual interpretation of digital aerial images or field measurements using tachymeters, are either inaccurate or expensive. Recently, the development of laser-scanning technology has opened new opportunities for tree mapping and attributes updating. For a detailed measurement and attributes update of urban trees, we tested the use of a multisource single-tree inventory (MS-STI) for heterogeneous urban forest conditions. MS-STI requires an existing tree map as input information in addition to airborne laser-scanning (ALS) data. In our study, the tested input tree map was produced by terrestrial laser scanning (TLS) and by using a Global Navigation Satellite System (GNSS). Tree attributes were either measured from ALS or predicted by using metrics extracted from ALS data. Stem diameter-at-breast height (DBH) OPEN ACCESS Forests 2014, 5 1033 was predicted and compared to the field measures, and tree height and crown area were directly measured from ALS data at the two different urban-forest areas. The results indicate that MS-STI can be used for updating urban-forest attributes. The accuracies of DBH estimations were improved compared to the existing attribute information in the city of Helsinki’s urban-tree register. In addition, important attributes, such as tree height and crown dimensions, were extracted from ALS and added as attributes to the urban-tree register.


Introduction
Trees and woodlands are part of the urban environment, and they play an important role in urban areas [1].Urban woodlands can improve people's quality of life, but they also improve the urban environment.Trees can reduce pollution effects [2] and improve air quality [3] in urban areas.In relation to this, urban forests can also store industrial carbon emissions [4], and when vegetation is sustained and maintenance practices are focused upon (e.g., energy conservation or producing long-term carbon storages), urban vegetation can be seen as a carbon sink [5].
Urban forests also offer recreational environment for citizens.However, use level, user composition, and the temporal distribution of activity types through commuting and recreation use may differ depending on the number and closeness of settlements, business areas and schools [6].Urban forests can also be seen as a measure for strengthening social structures for interaction and for reducing social exclusiveness [7].Urban forests and woodlands can provide multiple values and benefits (e.g., [8,9]) including education, economy, and urban biodiversity, in addition to the aspects mentioned above.
In Finland, the City of Helsinki's Street and Park Division maintains a digital tree register (approximately 40,000 trees) that includes trees that are situated at road sides and some of the trees in the parks.The tree-register data includes information on the species, height, diameter-at-breast height (DBH), and location.The tree-register data are used in city and environmental planning, in locating old trees that are hazardous (for citizens), and in biodiversity monitoring.Trees in the register have been located on city-planning maps.The City of Helsinki's Street and Park Division is also interested in expanding the register to include trees in the parks that are not yet registered.The requirements for up-to-date tree data in city parks and forests are increasing, and an important question is how to keep the digital databases current for the various applications.Traditional updating procedures, such as visual interpretation of digital aerial images or field measurements using tachymeters, are either inaccurate or expensive.Hence, the utilization of remote sensing data is appealing.Remote sensing data are in many cases collected automatically for other purposes and are therefore affordable for updating digital databases.As in many cities, this is also the situation in the City of Helsinki where aerial photographs and airborne laser scanning (ALS) data are collected for other urban planning purposes, such as mapping of buildings, roads, and other built objects.New opportunities for tree mapping and the updating of attributes have been opened because of the development of laser-scanning technology.
ALS is used in urban planning to create highly detailed digital surface models and eventually Digital City Models and virtual-city realities [10].ALS has been used by forest companies and governmental organizations to acquire up-to-date information about forest resources.ALS provides a geo-referenced-point cloud, which enables the calculation of digital terrain models (DTMs) and, digital surface models (DSMs) that correspond to treetops and three-dimensional (3-D) models of an object (e.g., canopy-height model (CHM), normalized DSM), which are the main products used for laser-assisted forest measurements.The two main approaches to derive forest information from ALS data are the area-based approach (ABA) [11] and individual tree detection (ITD) [12].With ALS it is possible to directly measure the forest structure, including canopy height (CH) and crown dimensions; hence, it is increasingly being used for forest inventories at different levels.It has been shown that ALS can be used for estimating a variety of forest attributes, including tree, plot, and stand-level estimates for tree height [12][13][14][15], stem volume [16][17][18][19], basal area (BA) [11,[20][21][22] and tree species [23][24][25][26][27].
Holopainen et al. (2013) [31] demonstrated that TLS and mobile laser scanning (MLS) can be used for producing accurate tree maps in urban forests.Based on the results presented in [31], ALS was selected as a means of updating Street and Park Division's tree register for the city of Helsinki [43].In [43] an ALS-based mapping system was developed to be used for urban-tree mapping.On the basis of these studies, a conclusion can be drawn that urban-tree mapping will be done automatically in the near future.The next phase is to study how to measure and update automatically important tree attributes by using an existing tree map, which is a basic product in the urban-tree inventory.One especially interesting option is a combination of ALS and TLS data in which ALS can be used for measuring tree height and TLS is used for providing not only the accurate position of all trees but also detailed information about the quality of the tree stem.This method is useful for updating information of urban trees and also for developing precision forestry [44] in managed forests.
Here we tested the use of a multisource single-tree inventory (MS-STI) where ALS and TLS data were combined for mapping the trees and measuring tree variables.Our objectives were to investigate estimation accuracy of tree variables, estimate stem distributions, and asses the prediction accuracy while also producing completely new tree attribute such as crown size and update height for all trees in the city of Helsinki's urban tree register.Stem distribution can be used for urban-forest planning and for visualizing heterogeneous urban woodlands.In the future, 3D-virtual forests could be an essential part of urban planning, and stem distribution is an offset for creating these models.

Study Area
The study area, Seurasaari (in Helsinki, Finland), is a popular outdoor recreation area located approximately 5 km from the Helsinki city center.It was made a public park in 1890 and quickly became a popular place for recreational activities.Seurasaari is a wooded island with rocks, hills, wetlands, and herb-rich forests covering about 46 ha.It receives hundreds of thousands of visitors per year.Our study area in Seurasaari comprised two parts, covering approximately 2.7 ha in total (Figure 1).The northern part is a well-managed urban park comprised mainly of widely separated old oaks and only grass as under-storey vegetation, while the southern part resembles more a natural unmanaged park forest with varying under-storey vegetation.The distribution of tree species was diverse and consisted of 11 different species (Table 1), which describes the heterogeneity of the research area.In the area, there is a dense network of artificially constructed outdoor paths that can also be used by vehicles.

Field Measurements
A predefined TLS tree map was used to identify each tree.The DBH was determined for 389 trees.Steel callipers were used for the DBH measurements.The average DBH for the entire study area was 268 mm and varied between 31 and 482 mm.Descriptive statistics were also calculated for two separate areas: the northern part, which is a well-managed urban park forest, and the southern part, which is denser, unmanaged forest with varying under-storey vegetation.The average DBH was 371 mm in the northern part and 261 mm in the southern part.

Airborne Laser Scanning
The ALS data were acquired in 2011 with an Optech 3100 laser scanner (Optech Inc., Vaughan, ON, Canada).The flying altitude was 400 m.The density of the pulses returned was approximately 10 points per m 2 , and the return type was recorded (first-of-many, single, intermediate, last).The ALS data were first classified as ground and non-ground points according to the method developed by Axelsson (2000) [45].Low-point classification was also used to improve the accuracy of the ground level.A DTM was developed from classified ground points, and laser heights above the ground (normalized height or canopy height) were calculated by subtracting the ground elevation from the laser measurements.

Terrestrial Laser Scanning
The TLS data were collected with a Leica HDS6100 TLS system (Leica Geosystems AG, Heerbrugg, Switzerland) in September 2010.The HDS6100 is a 690-nm phase-based continuous-wave laser scanner with a 360° × 310° field of view (FOV) upward, and its data acquisition rate is 508,000 points per second.The distance measurement accuracy is ±2 mm at a distance of 25 m.The circular-beam diameter at the exit and the beam divergence are 3 mm and 0.22 mrad, respectively.The point spacing is 6.3 mm at 10 m.Further detailed specifications are presented below (Table 2).TLS measurements for the study area were collected in multi-scan mode.The park areas were scanned as-is.Pre-scan preparations, e.g., removal of low vegetation, were not done, since it was not permitted in the city forest.The objective of the measurements was to obtain good point coverage.The data were collected in five to seven scans per group; in total, 52 scans were performed to cover the entire study area.We positioned the center-scan station and at least one reference-target ball of each scan group using a GNSS virtual-reference station (VRS) and a tachymeter.The center scans were placed so that the canopy layer did not block the GNSS satellite visibility.Subsequently, we transformed the scans into global coordinates according to the scanner and sphere-target locations measured.The reference targets were placed on the forest ground for point-cloud registration.The TLS point clouds within each group were co-registered, using reference targets.Co-registration was done with Leica's Cyclone software (Leica Geosystems), in which the reported overlap root-mean-squared error (RMSE) of the scans ranged between 2.3 and 6.3 cm.

TLS-Based Tree Map
Tree detection and location measurements were done manually in the present study using the 3D-environment of TerraScan (Terrassolid Ltd., Helsinki, Finland).Point clouds were processed in scan groups, and trees were searched within each group using visual interpretation.The process was performed through the following steps:

Extraction of Tree Height, Crown Area, and the Prediction Variables from the ALS Data
A raster canopy height model (CHM) was created from normalized ALS point-height data for tree-crown segmentation.The maximum ALS point height from first-of-many or single echoes was assigned to each CHM cell (resolution 0.5 m), and no-data cells were filled with the mean height value in the 3 × 3-neighborhood.Tree crowns were delineated using watershed segmentation described in [20,46,47].
Tree crown segments were linked to the field trees based on the TLS-measured location.If there were more than one tree within a segment, the segment was split equally.As segments from CHM represented tree crowns, areas for tree crowns were calculated as the area of each segment.Tree height was possible to update by using the maximum height of ALS points of each tree crown.
Statistical metrics describing tree crown density and tree height were calculated from the ALS point-data within the tree crown segment (Table 3).First, heights over 0.5 m were classified as coming from the vegetation (i.e., tree), and a vegetation-density ratio (vege) was then calculated as a ratio between vegetation heights and all heights within a tree crown.The following were other extracted metrics: maximum (H max ), average (H mean ), standard deviation (H std ), and coefficient of variation of heights (CV); height at 10%-90% percentiles (h 10 -h 90 ), and crown-cover density metrics as a proportion of heights below a certain relative tree height (p 10 -p 90 ).Only the heights belonging to vegetation (that is, heights over 0.5 m) were used in the calculation of these metrics.The metrics were extracted from the ALS data using all returns.

Estimation of the Tree Variables
Diameter-at-breast height (DBH) was predicted by means of ALS metrics using the nearest-neighbor (NN) approach.Tree variable measured in the field (i.e., DBH) was used as target observations, and tree-specific metrics derived from ALS data were used as predictors.Random Forest (RF, [48]) was applied in the NN search.Based upon the quality of results and desirable statistical characteristics (i.e., the capability to predict multiple response variables simultaneously, use a large number of predictors without the problem of over fitting, and evaluate accuracy with built-in functionality), the use of RF in NN estimation of forest variables is increasingly common (e.g., [19,47,[49][50][51]).In [50] and [51] it was demonstrated that the RF method is more robust and flexible for forest variable prediction when compared to other NN distance measures, such as Euclidian distance, Mahalanobis distance, or Canonical Correlation Analysis.In the RF method, several regression trees are generated by drawing a replacement from two-thirds of the data for training and one-third for testing for each tree.The samples that are not included for training are called out-of-bag samples, and they can act as a testing set in the approach.The measure of nearness in RF is defined based on the observational probability of ending up in the same terminal node in classification.Predictor's importance can be assessed in RF by using the predictor's scaled-importance values.To measure the importance after the RF run, the values of the predictor are permuted among the training data, and the out-of-bag error is again computed on this blended data set.The importance score for the predictor is then computed by averaging the difference in the out-of-bag error before and after the permutation over all trees.The score is normalized by the standard deviation of these differences.The R statistical-computing environment [52] and yaImpute library [53] were applied in the RF predictions.The yaImpute library is tailored to NN forest-attribute estimation.
In the present study, 1200 regression trees were generated, and the square root of the number of predictor variables was picked randomly at the nodes of each regression tree.Randomness was taken into account by running the RF method 100 times.The final result was the average of these runs.The number of neighbors differed from 1 to 5.
Prior to the modeling, the RF was used to reduce the number of the predictor variables.A step-wise looping procedure was used to iterate RF, discarding the least important of the candidate variables at each iteration, based on the variable importance, until only a single predictor variable remained.RMSEs were calculated for each predictor variable combination and analyzed before the final modeling.

Accuracy Assessment at the Tree and Stand Level
In this study, predictors and target observations were available for all of the trees.Therefore, the accuracy of the predicted variables at the tree level was evaluated by calculating bias (Equation ( 1)) and RMSE (Equation ( 2)) using out-of-the-bag samples.The relative bias and RMSE were calculated according to the sampled mean of the variable in question.The accuracies of the tree-level measurements were evaluated by calculating the bias and RMSE: (1) (2) where n is the number of observations, y i the value estimated from the field data for observation i, and the predicted value for observation i.
Stem-diameter distributions were compiled from the tree-level predictions and compared to the field measurements for both parts of the study area separately.The predicted stem-diameter distributions were evaluated by the error index (EI) introduced in [54]: (3) where f i and i are the numbers of stems in class i to be compared, k is the number of classes or bins, and w i is the weight of class i.
In addition, an alternative error index was calculated following [55], who used relative frequencies: where f i is the true and i is the predicted stem number in class i, k is the number of classes or bins, and N is the true and the predicted stem number of all diameter classes.A weight 0.5 was used to scale the error index between 0 and 1 where 0 means a perfect fit and 1 means that distributions do not overlap at all.The used bin size for both error indices was 2 cm.

Selecting Predictor Variables
Predictor variables that were used in further estimations were selected separately for forested (southern part) and park areas (northern part).There were in total 23 extracted features that were possible to be used in prediction.RMSEs were calculated for each predictor-variable combination, and a minimum number of predictors were chosen before the out-of-bag prediction accuracy started to increase notably.For the forested area (southern part), seven variables were selected as predictors, whereas for the park area (northern part), six variables were selected (Table 4).Depending on the used set of predictors, the relative-RMSE values varied between 25.0% and 33.8% in the forested area (Figure 3) and between 18.8% and 20.1% in the park area (Figure 4); thus, the prediction accuracy was not oversensitive to the number of predictor variables that were used.

Prediction Accuracy of Diameter-at-Breast Height
Prediction accuracy for DBH varied between the southern (forest) and northern (park) parts of the study area.The increment of the number of k decreased the relative RMSEs in the forested area.For the park area, the relative RMSE increased when more than one neighbor were used.The relative RMSEs for the park area were smaller (from 10.7% to 19.1%) than for the forested area (from 26.3% to 29.1%) (Table 5).On the other hand, biases were bigger for the park area (from −4.2% to 1.8%) than they were for the forested area (from −2.5% to −0.4%).
The accuracy of the DBH measurements is highly affected by stem form, because tree stem form is not fully circular (Figure 5).Especially stem forms of urban trees can differ greatly from the circle form.This affects both the field measured and the predicted DBH accuracy.

Comparison of Stem-Distribution Series
Stem-diameter distributions were estimated separately for the northern (park area) and southern (forest area) parts of the study area.The main interest was in the forested area, since it could be better compared with results from managed forests.For both areas, the extreme bins of the DBH classes were omitted in the predicted stem-diameter distributions when the number of k increased (Figures 6 and 7).The number of trees in large DBH classes was especially difficult to predict.The Reynolds' error indices for the predicted stem-diameter distributions varied from 70 to 152 for the forested area (southern part) and from 6 to 26 for the park area (northern part) (Table 6).Reynolds' EI cannot be compared directly to other studies because they show the absolute value of the number of trees that do not match to the same DBH bins when comparing two distributions.The relative EI values showed better fit in the predicted stem-diameter distributions for the forested area than for the park area.

Extraction of Tree Height and Crown Area
It was possible to update the tree height by using the maximum height of ALS points of each tree crown.For the forested area (southern part of the study area), the tree height varied between 4.5 m and 29.2 m with a mean of 19.7 m and a standard deviation 5.3 m (Figure 8).The respective values for the park area (northern part of the study area) were 7.8 m, 26.0 m, 19.4 m, and 5.0 m (Figure 9).The tree crown area was computed from tree-crown segments and it varied between 1.1 m 2 and 86.5 m 2 in the forested area and between 2.5 m 2 and 95.2 m 2 for the park area.The mean and standard deviation for the forested area were 22.4 m 2 and 15.9 m 2 , respectively.For the park area respective values were 34.

Discussion
Updating information of urban-tree characteristics has been challenging due to the high cost of traditional updating methods.ALS cost efficiently provides 3D information about urban trees, and this information can be used for planning urban-forest management practices and predicting the risk of forest damages.In this study, MS-STI was tested in two different urban-forest environments.In addition to ALS data, MS-STI requires an existing tree map as input information for a tree-attribute update.
Recent studies on single-tree inventory (e.g., [56][57][58]) have concentrated on managed forests, and the results have been promising due to conditions that are favorable for STI (mature stand, single-tree-species stand).Depending on forest conditions, the results of STI have varied considerably [47,[59][60][61]).A tree map is a challenging prerequisite for the MS-STI analysis, but information from existing tree registers maintained by the city could be used as input for MS-STI.On the other hand, mobile mapping with imaging and laser sensors are also used more often for various surveying actions and could be used for tree mapping as well [31].The tested-input tree map was produced by TLS and used GNSS with an expected accuracy of below 0.1 m.Similar location accuracies can be obtained using 2D-or 3D-laser scanning [31,33,[62][63][64][65][66][67].Detailed tree-level information could be used for optimizing management practices, spatial modeling of the growth, and localizing potentially hazardous trees for residents or infrastructure.
Predictors for tree-quality attributes were extracted from ALS data for mapped-tree locations.Stem distribution was compiled by summing single-tree measures.The accuracy of the MS-STI was validated using field measures (DBH).It should be noted that comparing results of this kind of study to other studies is only suggestive and highly data dependent.Forest conditions in this study were very heterogeneous and differed considerably from managed, one-tree-species forests; hence, results from other STI studies that have been conducted in managed forests are not fully comparable with our results.After this reminder, we can point out that our results are promising, and for the park area (northern part of the study area) they are on the same level as results obtained earlier using STI methods [47,[56][57][58]68].For the forested area (southern part of the study area) that was more heterogeneous than the northern part, the RMSEs were not as accurate as results obtained with earlier studies, but the difference was marginal.By enlarging the number of neighbors, RMSEs for the forested area did get lower, but for the park area, on the other hand, the RMSE increased.Also, enlarging the number of neighbors decreased the goodness of fit of the estimated stem-diameter distributions.In general, the obtained stem-diameter distributions for the forested area were similar to the ones in [55] and [69] in which a comparable error index was used.In [69] the STI technique and tree-list imputation method were used, and error indices varied from 0.33 to 0.40 for all the trees, and the species-specific error indices varied from 0.9 to 0.23.Error indices we obtained with MS-STI varied from 0.10 to 0.21 for the forested area and from 0.11 to 0.48 for the park area.In [55] ABA was used to predict species-specific stem-diameter distributions, and error indices varied from 0.21 to 0.34.Relative EI values for the forested area were notably lower than the ones obtained from the managed forest for all tree species reported in [69].For the park area, on the other hand, the relative EI values were similar to those for all tree species reported in [69], with the exception of using only one neighbor when the EI was lower.With tree-species-specific EI values, Vauhkonen et al. (2013) [69] obtained better results.
Tree-height and crown-size information were extracted from the ALS data.Tree-height measurement accuracy from the ALS data is expected to be ±1 m [43,47,57,70,71] and that is close to the accuracy of clinometer measurements [72] traditionally used for measuring tree heights in the field.In Helsinki, the tree register does not contain tree-height information for park trees, because height measurements have been too laborious.In urban areas, tree-crown size is required for the planning and allocating of management practices.For example, this information can be used in locating trees that are interfering lamp posts, buildings, or drivers visibility at road junctions.
For urban tree-attribute updates MS-STI is an applicable method.Once the tree map is produced, there are various possibilities for updating the tree-attribute data.One intriguing application for MS-STI in green urban environments is a 3D-virtual forest, which can be used for visualizing the trees and modeling the effects of the alternative management options.This can be used to support decision making in green urban environments.In forthcoming studies, TLS-and MLS-based tree mapping and tree-species detection should be developed further with MS-STI.Required accuracy of the tree maps also requires further investigations.

Conclusions
This study showed the high potential of ALS and TLS measurements in very heterogeneous urban-forest environments.Here we tested a multisource single-tree inventory (MS-STI) method.In MS-STI, a tree map is produced by TLS and used as an input for ALS-based tree-attribute updating.We conclude that MS-STI is capable of providing realistic stem distributions and accurate tree-attribute update in an urban environment.Thus, the approach can be used for acquiring urban tree information and producing tree variables, such as tree height and crown size, for existing tree registers maintained by cities.
(1) the point cloud of each scan group was imported into TerraScan by thinning the point cloud by 50%; (2) points from an approximately 1.3-m height (between 1.25 m and 1.35 m to be exact) were classified into a horizontal -slice‖ for identifying tree trunks; (3) tree trunks were detected and marked within the slice (see Figure2); and (4) location and DBH information were recorded for all the trees.

Figure 2 .
Figure 2. Example of the tree-detection method from the TLS point clouds.

Figure 3 .
Figure 3.Effect of number of predictors to DBH estimation accuracy in and selection of ALS features for the forested southern part of the study area.

Figure 4 .
Figure 4. Effect of number of predictors to DBH estimation accuracy in and selection of ALS features for the park (northern part of the study area).

Figure 5 .
Figure 5. Two examples of measured DBH from TLS-point clouds from two different directions.

Figure 6 .
Figure 6.True stem-diameter distribution (top left) and predicted stem-diameter distributions with different k values for the forested southern part of the study area.

Figure 7 .
Figure 7. True stem-diameter distribution (top left) and predicted stem-diameter distributions with different k values for the park area (northern part of the study area).

Figure 8 .
Figure 8. Tree-height distribution using maximum height of ALS points of tree crowns for the forested area (southern part of the study area).

Figure 9 .
Figure 9. Tree-height distribution using maximum height of ALS points of tree crowns for the park area (northern part of the study area).

Table 1 .
Relative tree-species distribution in the study area.

Table 2 .
Leica HDS6100 TLS system and specifications.

Table 3 .
Statistics, such as minimum (min), maximum (max), range, mean, and standard deviation (Std) calculated from the extracted predictor variables.

Table 4 .
Selected predictor variables for both parts of the study area in order of the predictor importance.

Table 5 .
Accuracy in DBH prediction.

Table 6 .
Reynolds'error indices for the predicted stem-diameter distributions using a bin size of 2 cm.