Automatic detection of harvested trees and determination of forest growth using airborne laser scanning

Remote Sensing of Environment 90 (2004) 451 – 462 www.elsevier.com/locate/rse

Automatic detection of harvested trees and determination of forest growth using airborne laser scanning Xiaowei Yu a, Juha Hyyppa¨ a,*, Harri Kaartinen a, Matti Maltamo b a

Department of Remote Sensing and Photogrammetry, Finnish Geodetic Institute, P.O. Box 15, FIN-02431 Masala, Finland b Faculty of Forestry, University of Joensuu, Finland Received 19 June 2003; received in revised form 12 February 2004; accepted 14 February 2004

Abstract This paper demonstrates the applicability of small footprint, high sampling density airborne laser scanners for boreal forest change detection, i.e. the estimation of forest growth and monitoring of harvested trees. Two laser acquisitions were carried out on a test site using a Toposys-1 laser scanner. Three-dimensional canopy height models were calculated for both data sets using raster-based algorithms. Objectoriented algorithms were developed for detecting harvested and fallen trees, and for measuring forest growth at plot and stand levels. Out of 83 field-checked harvested trees, 61 could be automatically and correctly detected. All mature harvested trees were detected; it was mainly the smaller trees that were not. Forest growth was demonstrated at plot and stand levels using an object-oriented tree-to-tree matching algorithm and statistical analysis. The precision of the estimated growth, based on field checking or statistical analysis, was about 5 cm at stand level and about 10 – 15 cm at plot level. The authors expect that the methods may be feasible in large area forest inventories that make use of permanent sample plots. Together with methods for detecting individual sample trees, the methods described may be used to replace a large number of permanent plots with laser scanning techniques. D 2004 Elsevier Inc. All rights reserved.

1. Introduction Forests are living ecosystems, influenced by continuous natural and anthropogenic processes. Natural forest processes include annual growth, mortality and natural disasters, whereas cutting and cultivation present typical anthropogenic actions. Forest changes can be found either by detecting actual forest and land use change or by executing an inventory twice on the same area (Varjo & Mery, 2001). Remote-sensing-based methods utilize the former approach while national forest inventories typically use the latter. For example, the National Forest Inventory of Finland (Tomppo, 1991, 1997) utilizes permanent sample plots to monitor tree recruitment, growth, health and mortality. Tree coordinates on permanent plots are registered to identify them during the next inventory.

Abbreviations: Trees; Forest; Laser. * Corresponding author. Tel.: +358-9-29555305; fax: +358-929555200. E-mail address: [email protected] (J. Hyyppa¨). 0034-4257/$ - see front matter D 2004 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2004.02.001

Remote sensing techniques have been widely studied during the last decades to support or replace conventional forest inventories. Manual photointerpretation has been widely used for forest planning in Europe since the 1960s (Avery, 1966). The development of more automated forest inventory techniques has been more difficult, and the low success rate of satellite data applications has been due to high information requirements (e.g. the tolerated error in a standwise inventory in Finland is 15% for the main stand parameters). According to Tokola and Heikkila¨ (1997), reliable estimates of stand parameters can be obtained from Landsat images only for forest areas larger than 100 ha. The concept of using an airborne laser system to measure forest biomass evolved from oceanographic applications of lidar (LIght Detection And Ranging). This was made possible by the development of the laser systems. Nelson et al. (1984) recommended the use of the laser-derived stand profiles for the retrieval of stand characteristics. Subsequently, laser measurements have been studied in the estimation of tree height, stem volume and biomass (e.g. Nelson et al., 1988). Nilsson (1990) demonstrated that the data collected with a laser mounted

452

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

on a boomtruck correlates with volume changes, such as thinnings. The development of high-resolution aerial photographs has made it possible to use pattern recognition methods to detect individual trees (Brandtberg & Walter, 1998; Dralle & Rudemo, 1996; Gougeon, 1995) and the development of laser scanning techniques has made it possible to measure individual trees in 3D (Brandtberg, 1999; Hyyppa¨ & Inkinen, 1999; Hyyppa¨ et al., 2001a; Hyyppa¨ et al., 2001b; Lim et al., 2001; Næsset & Økland, 2002; Persson et al., 2002; Popescu et al., 2002; St-Onge, 2000). Hyyppa¨ and Inkinen (1999) demonstrated the possibility to measure single tree parameters (height, crown width, tree species). According to Persson et al. (2002), individual trees can be detected with up to 70% accuracy and tree height can be measured with an accuracy better than 1 m. In comparison to single tree-based estimation with orthoimages, laserbased crown delineation is a more robust technique, since it measures the geometrical properties of trees directly (Hyyppa¨ & Hyyppa¨, 2001). The detection of changes in forested areas using remotely sensed imagery has so far not been accurate enough to meet the assessment requirements, especially regarding slight or moderate changes such as thinning or forest damage (Ha¨me, 1991; Olsson, 1994; Varjo, 1996). Advanced techniques such as segmentation may improve change detection (Pekkarinen, 2002). Laser scanning can also be applied to monitor harvested and fallen trees, forest growth or changes in foliage or needle mass. Change detection methods are traditionally pixel-based, but it is possible to carry out an object-oriented change detection analysis by using object-oriented image analysis and postclassification comparison. A suitable object for forest inventories is an individual tree. This paper demonstrates for the first time the usefulness of small footprint, high sampling density airborne laser scanners for the detection of harvested trees and the estimation of tree growth. The

methods and applications were developed using laser scanner data acquired between autumn 1998 and early summer 2000.

2. Methods 2.1. Test site The test site used is located in Kalkkinen, 130 km north of Helsinki, in southern Finland (Fig. 1). A total of approximately 100 ha of state and private forests were selected for the study. Situated about 110 m above sea level, the test site is dominated by small hills and the main tree species are Picea abies (Norway spruce) (49%), Pinus sylvestris L. (Scots pine) (35%), Betula verrucosa and Betula pubescens (silver and downy birches) (11%). During the last decades, in most of the state forests hardly any silvicultural operations have been carried out. Similarly, in the private forests only some cuttings have been made. Twenty stands in the test site were selected for general demonstration of growth measurements. The median size of these stands was 1.2 ha and their delineation was based on a previous operational forest inventory. Forest inventory statistics are shown in Table 1. The same test site has also been used for the comparison of various remote sensing data sources for forest inventory (Hyyppa¨ & Engdahl, 1999; Hyyppa¨ & Hyyppa¨, 1999; Hyyppa¨ et al., 2000). 2.2. Laser acquisitions Laser scanning is a method based on distance measurements (using laser range measurements) and the precise orientation of these measurements (using inertia measurements) between a sensor (the position of which is known as (x1, y1, z1)) and a reflecting object (the position of

Fig. 1. Location of test site shown on the map of southern Finland.

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462 Table 1 Descriptive statistics of selected 20 stands

Median Mean S.D. Min Max

2.3. Field inventories

Size (ha)

Age (years)

Basal area (m2/ha)

Mean diameter (cm)

Mean height (m)

Volume (m3/ha)

1.2 1.96 2.24 0.3 10.1

82 70.4 29 14 105

25.9 26.6 7.8 11.9 37.8

25.2 22.2 7.9 4.7 29

20.8 18.4 6.4 4.5 24.1

254.3 237.6 110.2 31.3 395.1

which is to be defined) giving the co-ordinates of the object (x2, y2, z2). For the purposes of forest inventories, the sampling density (laser hits per unit area), scan angle, and profile information require careful validation. A high sampling density is needed to detect individual tree crowns and a steep scan angle increases the number of ground hits. Test flights (TopoSys, 1996) have shown that at scan angles of more than 10j off-nadir shadowing increases heavily, i.e. the number of ground hits decreases and gaps in the digital terrain model (DTM) occur more frequently. The first and the last pulse, i.e. the first and last echoes of distributed targets such as the forest, are typically both acquired during the same flight. The Toposys-1 laser scanner was selected for this demonstration, due to its high sampling rate and steep scan angle. The laser data were acquired on 2– 3 September 1998 and 15 June 2000. Three DGPS receivers were employed to record the carrying platform position: one on the aircraft, and two on the ground (the first as the base station, the second for backup). The test site was measured from an altitude of 400 m resulting in a nominal sampling density of about 10 measurements per square meter. This altitude was selected in order to provide the number of pulses needed to resolve individual trees. Table 2 presents the system parameters of the Toposys-1 sensor.

Table 2 TopoSys-1 laser scanner performance parameters Characteristics

Description of Performance(s)

Sensor

Pulse-modulated, TopoSys-1, see www.toposys.com 83,000 Hz 630 Hz F 7.1j 4. . .5 m 2 at 800 m 128 parallel shots (one of which is the reference) 200 m x, y < 1.0 m z < 0.15 m Class 1 by EN 60825 (eye-safe)

Laser pulse frequency Scan frequency Field of view Measurement density Number of shots per scan Swath width at 800 m Position accuracy Elevation accuracy (WGS84) Laser classification

453

2.3.1. Harvested trees Fallen or harvested trees were located in the field. Altogether 26 individual harvested trees, one fallen tree and 15 clusters consisting of a total of 56 young harvested trees were found. 2.3.2. Vertical growth Tree height and shoot elongation were measured in August 2002. Only pine shoots were measured, since annual shoot growth of other tree species is much more difficult to determine. Using a tacheometer, a theodolite or a glass fibre, first the total height of the trees was measured and then five consecutive shoots below the top of the tree. This gave the height of the tree after each annual growth period between 1997 and 2002. Measurements were concentrated on three plots (A, B and C) with an average size of 25  30 m for 91 trees (Fig. 1). For plot A, consisting of saplings, trees could not be identified on laser images and thus tree-to-tree comparison was not possible. For plots B and C, trees could be linked to laser detected trees using either the coordinates measured for the trees or a laser-derived canopy height image applied in the field. The descriptive information of the measurements is shown in Table 3. Additionally, a plotwise inventory, in which all trees were measured and located, had already been carried out in 2001 in the same test site (Vehmas, 2002). One of the difficulties in the comparison between the reference data and laser-derived values was the 2-year gap between the two laser acquisitions. In boreal conditions the annual growth period lasts approximately from the beginning of May to the end of August. Height growth differs between tree species. According to Kanninen et al. (1982), in Finland most height growth of Scots pine is completed by 15 June. 2.4. Laser data preprocessing for change detection To develop single tree-based methods for harvested tree detection and forest growth estimation, some preprocessing of the data was necessary, such as the calculation of canopy height models and segmentation of individual tree

Table 3 Descriptive statistics of field inventories for growth evaluation Parameters Number of trees Height in 2000 (m)

Growth (1998 – 2000) (m)

Plot

A

B

C

Average Min. Max. Average Min. Max.

63 6.4 3.41 8.44 1.12 0.57 1.45

15 17.32 9.44 24.08 0.45 0.23 0.69

13 15.75 7.41 22.51 0.52 0.16 0.85

454

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

crowns. Data from both acquisitions were processed separately using the same algorithms and approaches. 2.4.1. Creation of canopy height model (CHM) The canopy height model was computed as the difference between the digital surface model (DSM), representing the tops of the crowns, and the digital terrain model (DTM). The DSM of the crown was obtained by taking the highest value (z value) of all laser hits within each pixel (50 cm). The value for missing pixels was obtained using Delaunay triangulation and linear interpolation. Hits coming from within the crown may be valuable for biodiversity information but are not needed in the construction of canopy height models. To generate a DTM from laser scanner data, points that are reflected from non-ground objects, such as from trees and buildings, must be classified as non-ground hits. This was accom-

plished by a method developed at the Finnish Geodetic Institute (FGI) based on an algorithm of Ruppert et al. (2000). The general idea is to combine a high resolution DTM with a low resolution one in order to identify and remove the non-ground points. It is easy to filter out vegetation and gaps with a lower resolution (10 m). In this study, the set of resolutions used was 7, 5, 3, 2, 1 and 0.5 m. The procedure started with a coarse resolution (7 m). The DTM was generated from raw laser scanner data by finding the minima of all the laser measurements within each pixel. First, the DTM of 7- and 5-m resolutions (DTM7 and DTM5, respectively) were created. Second, the DTM of 5-m resolutions was created again by interpolating (resampling) from DTM7, resulting in DTM57. Finally, the non-ground points were identified by applying the predefined threshold to differences between DTM5 and DTM57. Removal of non-ground points was

Fig. 2. An example of the automatic detection and location of harvested trees, (a) canopy height model of the 1998 data, (b) canopy height model of the 2000 data, (c) difference image, (d) image after thresholding and filtering, (e) segmented image indicating individual harvested trees.

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

455

the tree crown shape and location of individual trees were determined. First, trees were located by looking at the local maxima in the low-pass filtered canopy height model. Then a watershed-type segmentation procedure was applied. Each segment was considered to present a single tree crown. Each of the 20 stands was processed separately to get an optimum result. 2.5. Automatic detection of harvested trees

Fig. 3. A flow diagram showing the steps for the automatic detection of harvested trees.

done by replacing the identified points of DTM5 with their values in DTM57. This combination was then repeated as often as was necessary to generate a DTM of the desired resolution (0.5 m). Because of the action of taking the minimum value for each DTM pixel, the resulting DTM is usually lower than the actual ground surface, especially in hilly areas. Thus, a final refinement was performed by comparing raw laser points with the corresponding values from the created DTM. The points that differed less than 0.5 m were assumed to be ground hits and used in the creation of the final DTM. If there was more than one point in one pixel, the mean value was taken. The accuracy of the DTM algorithm was tested by Ahokas et al. (2002), and in a hilly, forested environment an accuracy of 14 cm was obtained. The FGI algorithm is comparable in accuracy with the commercial TerraScan software (Ahokas et al., 2002), which is based on the method by Axelsson (1999). 2.4.2. Delineation of single trees There are several techniques for automatic identification of individual trees using high resolution imagery (Friedlaender & Koch, 2000; Persson, 2001; Pouliot et al., 2002; Quackenbush et al., 2000; Stiteler & Hopkins, 2000). Most of these were originally developed for delineating trees from high resolution aerial imagery. In this study, for 20 selected stands single tree-based segmentations were performed on canopy height model images derived from both the 1998 and the 2000 acquisitions using commercial software (Arboreal Forest Inventory Tools of Arbonaut). The algorithm used is described in Hyyppa¨ et al. (2001a), and the performance of the segmentation algorithm compared to two other algorithms (developed at Joanneum Research in Austria and the University of Freiburg in Germany) is described in Hyyppa¨ et al. (2001b). During the segmentation processes,

The method for detecting harvested trees was based on difference imaging: each pixel value in the image corresponding to the canopy height model of the year 2000 was subtracted from its corresponding pixel value in the image of the canopy height model of the year 1998. The resulting difference image represented the pixel-wise changes between the two dates. High positive differences are from harvested trees (Fig. 2). In this study, a method for automatic identification was developed (Fig. 3). First, a threshold value was applied to the difference image in order to distinguish major changes. Then, a morphological opening was performed to reduce noise-type fluctuation. Finally, based on the segmentation of the resulting image, the location and number of harvested trees was determined. 2.6. Estimation of tree height growth Trees in the boreal forest zone grow at a rate of 5 cm to 1m per year, depending on the tree species and size. According to Hyyppa¨ and Inkinen (1999), the height of individual trees can be determined with an accuracy of approximately 1 m. Therefore, it is extremely difficult to estimate single tree growth if measurements are taken with only a 2-year interval, as was done in this case. In

Fig. 4. The laser pulses hit the trees usually missing the tops of the trees.

456

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

Fig. 5. Tree to tree match results for stand 135; left: 1998 data, right: 2000 data. Segments marked with + were matched trees.

particular, the tip of the trees can be missed by the laser hits (Fig. 4). Practical forest operations require average growth estimation at plot or stand level, not at tree level. This can be obtained by calculating the mean difference of all tree heights taken at two acquisitions. However, if they are not considered in the analysis, activities which have occurred after the first acquisition, such as selective thinning or cutting, could lead to false interpretations and to systematic errors in interpretation. Therefore, a tree-totree matching algorithm was developed to locate the trees which existed at the time of both acquisitions. Single tree heights were extracted from the canopy height model by taking the highest value of each corresponding tree segment. The location of the trees was defined as the location corresponding to the highest value of the segments. If the locations of two segments, one from each acquisition, were within a specified distance (0.5 m), which we called threshold distance, then these two segments were considered a match. Fig. 5 presents an example of matching results for one selected stand. The laser-derived height growth of a tree was calculated by subtracting the height obtained in 1998 from the height measured in 2000 based on matching results. The laserderived growth on the stand or plot level was obtained by the mean value of the height differences for all matched trees. Outliers of matched trees (which resulted from tree-to-tree mismatching) were excluded by removing those trees whose height difference was greater than twice the standard deviation of differences from the mean. This was done iteratively until there was no significant change in the mean growth. 2.7. Changes in crown cover percentage In addition to growing in height, the crown also grows horizontally, leading to an increase in crown cover percent-

Table 4 Validation of harvested tree detection

Field-measured Laser-derived

Number of single cut trees

Number of trees and groups

27 32

56 trees in 15 groups 34 trees in 13 groups

age. The same tree-to-tree matching technique and statistical analysis were also used to analyse the change in crown cover percentage.

3. Results 3.1. Analysis of automatic detection of harvested trees A comparison of automatically detected harvested trees with field measurements showed that harvested trees were detected with a high accuracy (out of 83 trees, 61 were correctly detected). Only two small clusters of harvested trees (where the radius of crown areas was less than 1 m) were missed. The rest of the harvested trees or clusters were detectable, but the number of harvested trees recognised

Table 5 Statistics of height growth for stands Stand

Size (ha)

126 128 135 136 137 139 140 147 148 149 153 156 158 185 188 189 190 191 192 200

0.7 10.1 0.5 4.9 3.7 0.8 1.6 2.3 1.2 0.7 1.9 1.7 3 1.2 0.4 0.3 1.2 1.2 0.8 1

Fieldmeasured volume (m3/ha)

347.0 112.3 44.0 254.3 328.3 337.1 274.6 395.1 272.6 295.7 214.0 195.4 365.3 101.7 222.5 31.30 248.50

N

Laserderived height (m) in 2000

44 923 87 780 444 52 206 333 64 45 245 146 245 110 60 48 132 92 92 49

8.90 25.45 17.14 7.09 21.38 16.87 21.42 22.64 21.46 19.26 24.67 22.08 19.45 23.21 16.14 21.77 8.58 6.26 18.55 15.23

Laserderived mean growth (m)

S.D. (m)

S.E. (m)

0.66 0.10 0.48 0.60 0.22 0.41 0.20 0.04 0.04 0.01 0.09 0.05 0.16 0.09 0.11 0.20 0.58 0.75 0.21 0.67

0.22 0.38 0.40 0.56 0.32 0.56 0.49 0.35 0.35 0.28 0.26 0.30 0.40 0.38 0.50 0.42 0.59 0.38 0.29 0.38

0.03 0.01 0.04 0.02 0.01 0.08 0.03 0.02 0.04 0.04 0.02 0.02 0.03 0.04 0.06 0.06 0.05 0.04 0.03 0.05

N—the number of matched trees within the stand. S.D.—standard deviation of laser-derived growth. S.E.—standard error of the mean growth.

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

457

Table 6 Comparison of two tree selection methods for growth estimation

Stand Average height (m) of selected trees in 1998 Average height (m) of selected trees in 2000 Growth (m) (98 – 00) Fig. 6. Percentage of match for 20 stands plotted as a function of average tree height.

differed from the number obtained in the field measurements (Table 4). The differences in the number of detected trees were mainly due to mistakes in tree delineation. If the tree crown contained considerable height variations, a single tree was split into two segments. If the trees grew close to each other, several trees were merged into one segment. The most difficult thing was the recognition of small deciduous trees. Encouragingly, several partially harvested trees (where the large branches had been cut off) near a new power line were automatically recognised, even though they were neither circular in shape nor large in size. 3.2. Analysis of height growth using laser scanning Matched trees from 20 selected stands were used to calculate laser-derived growth at the stand level (Table 5). Typically, standard errors were less than 5 cm and in all cases less than 10 cm. However, due to the inaccuracy of the segmentation and small threshold value for matching, not all trees were matched. The percentage of correct matches for 20 selected stands ranged from 39% to 70% with a threshold distance of 0.5 m. The percentage of matches showed a

Fig. 7. Relation between percentage of match and threshold distance for one selected stand (137).

All trees selected

Tree-to-tree matching for selection

137 20.73

139 17.83

137 21.16

139 16.46

21.20

16.85

21.38

16.87

0.47

0.98

0.22

0.41

modest correlation with stand height (Fig. 6) (and, therefore, with stand age). In general, young stands tended to have a lower percentage of correct matches whereas old stands had a higher one. The percentage of matches rose with a rising threshold value (Fig. 7). By increasing the threshold value from 0.5 to 1 m, the percentage of matches increased from 52% to 80% for stand 137. However, while the percentage of matches increased, the number of mismatches also increased. The results implied that more accurate results are obtained with small tolerance values, even though the number of matched trees is smaller. Although some trees are missed when using a matching technique to estimate height growth, the use of matched trees in the analysis instead of all detected trees leads to a decrease in the systematic error due to cuttings (Table 6). In stand 137 no trees were harvested between the laser acquisitions. In stand 139 some trees were cut. The tree-to-tree matching algorithm distinguished the harvested trees and provided a reliable growth estimate of 41 cm, while the use of all detected trees resulted in a negative growth of almost 1 m. This was due to the removal of some large trees. Since larger trees have a greater chance to find a match than smaller ones, using tree-to-tree matching leads to a smaller growth estimate for stand 137. This also implies that growth would be better expressed as a function of tree height and species within the stand.

Fig. 8. Scatter plot showing relation between heights of matched trees in 1998 and 2000; the systematic difference gives the mean growth and the standard deviation describes the uncertainty in measuring single tree heights.

458

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

Fig. 9. Histogram of height growth for stand 137; the negative values and large growths for individual trees are due to errors in obtaining single tree heights.

After tree-to-tree matching, it was possible to calculate the mean of the height differences of matched trees and thus also the mean growth (Fig. 8). The standard deviation described the accuracy for single tree height estimation. The standard error gave the precision of the growth estimates at stand level. Due to the inaccuracy in the measurement of individual tree growth, the precision of the estimates was based on obtaining a large sample size (Fig. 9). Fig. 10 shows the growth estimation as a function of tree height with confidence limits. To study laser-derived growth as a function of tree species, information about the species of each tree was obtained from a plotwise measurement carried out in 2001 (Vehmas, 2002). Tree height was extracted from both canopy height models using a window area of 2.5  2.5 m. Growth was calculated separately for the two main species, i.e. pine and spruce (Fig. 11). On average, the laser-derived growth was 0.28 m for pine and 0.25 m for spruce. The results agree well with the statistics of forest growth in corresponding areas and in boreal forests more

Fig. 11. Laser-derived mean growth plotted as a function of tree height and species. Trees were divided into groups at 2-m height intervals. Height and growth are the average values of all trees within the intervals.

generally. Young pine tends to grow faster than spruce and old pine more slowly than spruce. Three plots were used for validating the laser-derived growth; this was done by comparing the field-measured values with the laser-derived ones. Since the laser beams do not necessarily hit the top of the trees (Fig. 4), it is probable that the laser underestimates the tree height. In order to verify the amount of underestimation, the laser-derived tree heights and field-measured tree heights were compared. Tree heights were systematically underestimated: by 67 cm in 2000 and 54 cm in 1998. This corresponded to 2– 3 years of growth. Of this, the elevation model overestimation (due to undervegetation) accounted for 1 year’s growth. In Table 7 growth was calculated for various combinations of growth periods. For plot A the 2-year growth between 1998 and 2000 (i.e. the growth periods of 1999

Fig. 10. Tree height growth with confidence intervals as a function of height for stand 137; trees were divided into groups at about 1-m height intervals. Height and growth are the average values of all trees within the intervals.

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

459

Table 7 Verification of height growth at plot level Plot

A

B

C

Field-measured growth (m)

Mean S. D. S. E. Mean S. D. S. E. Mean S. D. S. E.

Laser-derived growth (98 – 00) (m)

98 – 00

97 – 00

98 – 99

97 – 99

97 – 98

Before DTM compensation

After DTM compensation

1.120 0.193 0.024 0.447 0.129 0.032 0.521 0.201 0.056

1.739 0.276 0.035 0.671 0.204 0.053 0.856 0.391 0.108

0.541 0.097 0.012 0.220 0.071 0.018 0.276 0.106 0.029

1.160 0.180 0.023 0.442 0.158 0.041 0.611 0.302 0.084

0.619 0.109 0.014 0.220 0.089 0.023 0.335 0.216 0.060

0.862 0.569 0.104 0.384 0.418 0.112 0.038 0.589 0.170

1.139 0.664 0.121 0.445 0.389 0.104 0.452 0.540 0.156

S.D.—standard deviation. S.E.—standard error.

and 2000) was 1.12 m and between 1997 and 1999 (i.e. the growth periods of 1998 and 1999) it was 1.16 m. The laserderived growth was 86.2 cm. For plot B the corresponding field-measured values were 44.7 and 44.2 cm and the laserderived growth was 38.4 cm. For plot C, the figures were 52.1, 61.1 and 3.8, respectively. For three plots the change in growth between two growth periods (1998 – 2000 versus 1997 –1999) was smaller (by 4, 0.5 and 9.0 cm) than the errors of the growth estimation. Thus, this effect was neglected in the subsequent analysis. As can be seen from Table 7, the laser-derived data showed an underestimation of growth for all three plots. By investigating the data, we found that when comparing the

Table 8 Statistics of laser-derived crown cover change at stand level Stand

N

Laserderived height (m) in 2000

126 128 135 136 137 139 140 147 148 149 153 156 158 185 188 189 190 191 192 200

53 807 89 774 475 63 179 296 39 45 275 151 223 105 49 41 135 93 102 45

8.90 25.45 17.14 7.09 21.38 16.87 21.42 22.64 21.46 19.26 24.67 22.08 19.45 23.21 16.14 21.77 8.58 6.26 18.55 15.23

Change in crown cover for matched tree (%) 7.12 1.99 6.86 5.02 3.77 2.53 2.22 2.95 0.09 0.25 1.47 0.89 0.02 3.5 7.56 2.16 0.54 5.55 7.42 2.14

S.D. (m2)

S.E. (m2)

7.29 3.83 2.87 4.4 3.18 6.07 2.58 3.74 0.63 2 4.07 2.76 2.67 4.36 3.15 2.36 3.04 6.55 5.11 8.68

1 0.13 0.3 0.16 0.15 0.77 0.19 0.22 0.1 0.3 0.25 0.22 0.18 0.43 0.45 0.37 0.26 0.68 0.51 1.29

N—the number of matched trees. S.D.—standard deviation. S.E.—standard error.

DTMs derived from the two acquisitions, both systematic and random errors were found. To remove these errors from the analysis, DTM compensation was performed. This ensured that the same DTM was used in the estimation of height growth, which is important since the terrain is unchanged between the acquisitions. After DTM compensation, the laser-derived growth for plots A, B and C were 1.14 m, 44.5 cm and 45.2 cm, respectively (Table 7), and the corresponding standard errors were 12, 10 and 15 cm, which was reasonably well in agreement with the field measurements. 3.3. Analysis of crown cover percentage The same 20 stands were used to study horizontal growth (Table 8). As expected, the results showed a trend confirming that young trees also grow horizontally more rapidly than old trees. Crown cover at stand level changed between 0.54% and 7.56%. The growth of a tree crown area can be seen more clearly using difference imaging, where the difference of the canopy height models of the two dates is presented (Fig. 12). It was not possible to validate the horizontal growth because field measurements were not available.

Fig. 12. Difference image of maximum canopy height model between 1998 and 2000 data showing the horizontal growth of crowns. Maximum canopy height model was created by assigning the tree height value to all elements of the tree segment. Image displayed in black and white for better visualization. Horizontal growth of trees or cluster of trees is shown in white.

460

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

4. Discussion One critical aspect that could strongly affect the results is the segmentation of a single tree. The algorithm applied tends to merge two or more trees into one segment rather than split one tree into two separate segments, which was also noticed by Hyyppa¨ et al. (2001b). Presently, most of the algorithms used for the segmentation of laser scanning data are developed for 2D images and thus do not take full advantage of the 3D information of laser scanner data. Tree-to-tree matching is directly influenced both by single tree delineation and by the matching algorithm. In this study, the matching rate for 20 stands varied significantly. The lower the matching rate is, the fewer trees can be monitored automatically. Therefore an improvement in treeto-tree matching is important for obtaining an unbiased estimation of tree growth at plot and stand levels. The algorithm applied in this study was a simple distance-based algorithm for the location of the tree. Improvements in the segmentation will improve the matching significantly. In the future the trees should also be matched using the original point clouds produced by the laser. Although laser data underestimate tree heights (for details see Maltamo et al., forthcoming), laser-derived growth seems to be reasonably accurate at the stand level and if at least 100 trees can be matched for a stand, its precision is 5 cm or better. However, calculating laserderived growth for individual trees does not seem to be an easy task. A significant improvement could be made if the laser flights were flown with similar flight tracks, allowing the detection of trees with the same geometry. The results of this study can be compared to earlier studies concerning measurement and modelling of the growth of tree height (Hynynen et al., 2002; Pa¨ivinen et al., 1992). However, these studies deal with single trees. Pa¨ivinen et al. (1992) studied the measurement of growth of tree height and for a 5-year period of growth the random errors varied between 43 and 77 cm in different experiments. The errors were slightly smaller for Scots pine than for Norway spruce. In a study by Hynynen et al. (2002) predicting the growth of tree height for a 5-year period, the errors of the height growth models were 37.5 cm for Scots pine and 51.2 cm for Norway spruce. It should be remembered that the result of forest growth estimation based on tree-to-tree matching does not correspond to the average growth of a stand, it corresponds to the average growth of the trees visible and sampled with the laser scanner. Since the laser does not recognize all smaller trees, it is obvious that growth estimates are weighted by the tallest trees. This impact may be minor because in any one stand most of the volumetric growth is concentrated on the dominant trees. This phenomenon is also dependent on the stand structure. If multi-layered tree canopies are present, the concept of growth in a stand may be a complex one and the method used in this study may not give realistic growth figures for the whole stand.

However, managed boreal forests are usually single-layered. Since the laser-based growth can be estimated as a function of tree height within the stand and within the given confidence limits, as demonstrated in Fig. 11, the effects of weighting by the tallest trees can be corrected if the height distribution of the trees is known. The impact of the DTM on laser-derived growth was significant in all cases except one (plot B). However, the growth is usually defined over a longer period than in this study. In boreal conditions, typically a 5-year period is used both in presenting the average level of growth and also in models predicting growth (e.g. Hynynen et al., 2002). The average growth across such a period is considerable enough to make the errors arising from the DTM insignificant and therefore DTM compensation is not required in the growth estimation. However, the errors of the DTM need to be more carefully analyzed in the future. Our results implied that the underestimation of tree heights (with laser scanning) is actually partly caused by the overestimation of the DTM (due to undervegetation). Leckie et al. (2003) concluded that they had complete coverage of the entire area by a laser signal and, therefore, it was not likely that consistent underestimation of tree height would be due to the laser coverage not hitting the tops of trees. Rather, it was assumed to be related to a lack of lidar reflection off the top of the tree or the threshold at which the sensor detects a return signal, and partly also to the overestimation of the DTM due to the undergrowth. In addition to the DTM and changes in the forest, there are other factors that may have an impact on the laserderived height growth. Weather conditions, such as strong winds, can cause displacements of tree crowns. The direct georeferencing technique can also lead to small planimetric displacements. In hilly areas, measuring the same target from different directions can lead to differences. The processing of the laser scanner data requires that both data acquisitions should be processed with the same algorithms to avoid any algorithm-related change. The approach presented in this study offers a possibility to at least partially replace permanent sample plots in forest inventories. It is also possible to effectively examine the local variability of forest growth since laser scanning produces estimates not just for a small sample plot but for the whole area covered. Such growth data could also be utilized as reference material in non-parametric growth models. These have been found to be reliable when constructing local growth models (Sironen et al., 2003). Presently, the availability of laser data is improving significantly each year and the costs are steadily decreasing due to the acceptance of a new system with a higher sampling density and a higher flight altitude. Most of the presently available laser systems can detect individual trees. Present costs for laser surveys are highly dependent on the size and shape of the test site, the average cost being o2– 5 per hectare. In traditional forest inventories, the costs for

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

permanent sample plots are normally o100 per plot. The most economic use of laser scanning in forestry is to apply it to strip-based sampling, since long strips are economic to fly. Thus, large-area forest inventories using permanent or non-permanent sample plots are perhaps the most feasible operative applications for laser scanning at the individual tree level. The presented technique is not a cost-effective one when 100% coverage of the forest is needed, but it is feasible as a sampling technique. The technique presented is also economically feasible for monitoring trees in city parks.

5. Summary This paper is the first to present the detection of harvested trees using laser scanner data. Methods were developed for the automatic detection of such trees using difference imaging with canopy height models. Out of 83 field-checked harvested trees, 61 trees were automatically detected. All mature harvested trees were detected. The data did not allow the monitoring of some of the small trees in a reliable manner. This is also the first time forest growth estimation using laser scanner data has been demonstrated. Automatic methods to obtain growth statistics at plot and stand levels were developed on the basis of tree-to-tree matching. Individual tree change detection was not performed in a reliable manner. Since the stand level verification of height growth is extremely difficult, the analysis was based on statistical evaluation, i.e. by calculating the standard error of the mean to represent the precision of the growth estimate. Additionally, 91 trees were measured to give growth information at the plot level for three plots. The obtained standard error of the mean agreed well with the field measurements. Growth estimates between the data acquisitions did not necessarily represent the growth between these dates, since the laser pulse does not hit the tip of the tree, at least in young stands. Therefore, the obtained growth estimate may present, for example, the growth between the acquisition dates less 1 year. The precision of the growth estimated, based on field checking or statistical analysis, was typically about 5 cm at stand level and about 10 –15 cm at plot level. The authors expect the methods developed to be feasible in large area forest inventories where permanent sample plots are used to monitor individual sample trees and their growth at the plot level. The presented techniques may well allow the reduction of the number of permanent plots.

Acknowledgements The financial support by the Academy of Finland for the projects ‘‘Development of advanced retrieval algorithms and methods for laser scanning’’ (2001 – 2003) and ‘‘The usability of single tree laser scanning in forest planning’’ (2002 –2004) is gratefully acknowledged.

461

References Ahokas, E., Kaartinen, H., Matikainen, L., Hyyppa¨, J., & Hyyppa¨, H. (2002). Accuracy of high-pulse laser scanners for digital target models. In: Observing our environment from space. New solutions for a new millennium. Proceedings of the 21st EARSeL Symposium, Paris, 14 – 16 May, 2001 ( pp. 175 – 178). Balkema Publishers. Avery, T. E. (1966). Foresters Guide to Aerial Photo Interpretation. Forest Service Handbook No. 308. Axelsson, P. (1999). Processing of laser scanner data—algorithms and applications. ISPRS Journal of Photogrammetry and Remote Sensing, 54, 138 – 147. Brandtberg, T. (1999). Automatic individual tree-based analysis of high spatial resolution remotely sensed data. Dissertation, Swedish University of Agricultural Sciences, Uppsala, 47 pp. Brandtberg, T., & Walter, F. (1998). Automated delineation of individual tree crowns in high spatial resolution aerial images by multi-scale analysis. Machine Vision and Applications, 11, 64 – 73. Dralle, K., & Rudemo, M. (1996). Stem number estimation by kernel smoothing in aerial photos. Canadian Journal of Forest Research, 26, 1228 – 1236. Friedlaender, H., & Koch, B. (2000). First experience in the application of laserscanner data for the assessment of vertical and horizontal forest structures. International Archives of Photogrammetry and Remote Sensing, XXXIII, 693 – 700 (Part B7, Amsterdam). Gougeon, F. A. (1995). A crown-following approach to the automatic delineation of individual tree crowns in high spatial resolution aerial images. Canadian Journal of Remote Sensing, 21(3), 274 – 284. Ha¨me, T. (1991). Spectral interpretation of changes in forest using satellite scanner images. Dissertation, University of Helsinki, 111 pp. Hynynen, J., Ojansuu, R., Ho¨kka¨, H., Siipilehto, J., Salminen, H., & Haapala, P. (2002). Models for predicting stand development in MELAsystem. Finnish Forest Research Institute Paper, 835 (116 pp.). Hyyppa¨, H., & Hyyppa¨, J. (1999). Comparing the accuracy of laser scanner with other optical remote sensing data sources for stand attributes retrieval. The Photogrammetric Journal of Finland, 16, 5 – 15. Hyyppa¨, J., & Engdahl, M. (1999). Verification of the capability of repeatpass ERS-1/2 SAR interferometry to provide digital elevation models and the impact of the tree height and canopy colure on SAR-derived terrain height in boreal forests. The Photogrammetric Journal of Finland, 16, 16 – 26. Hyyppa¨, J., & Hyyppa¨, H. (Eds.) (2001). HIGH-SCAN—Assessing forest stand attributes by integrated use of high-resolution satellite imagery and laserscanner. Contract No. ENV4-CT98-0747 of European Commission. Final Report. September 2001, 81 pp. Hyyppa¨, J., Hyyppa¨, H., Inkinen, M., Engdahl, M., Linko, S., & Zhu, Y. -H. (2000). Accuracy comparison of various remote sensing data sources in the retrieval of forest stand attributes. Forest Ecology and Management, 128, 109 – 120. Hyyppa¨, J., & Inkinen, M. (1999). Detecting and estimating attributes for single trees using laser scanner. The Photogrammetric Journal of Finland, 16, 27 – 42. Hyyppa¨, J., Kelle, O., Lehikoinen, M., & Inkinen, M. (2001a). A segmentation-based method to retrieve stem volume estimates from 3-dimensional tree height models produced by laser scanner. IEEE Transactions on Geoscience and Remote Sensing, 39, 969 – 975. Hyyppa¨, J., Schardt, M., Haggre´n, H., Koch, B., Lohr, U., Scherrer, H. U., Paananen, R., Luukkonen, H., Ziegler, M., Hyyppa¨, H., Pyysalo, U., Friedla¨nder, H., Uuttera, J., Wagner, S., Inkinen, M., Wimmer, A., Kukko, A., Ahokas, A., & Karjalainen, M. (2001b). HIGH-SCAN: The first European-wide attempt to derive single-tree information from laserscanner data. The Photogrammetric Journal of Finland, 18, 43 – 53. Kanninen, M., Hari, P., & Kelloma¨ki, S. (1982). A dynamic model for above ground growth of dry matter production in a forest community. Journal of Applied Ecology, 19, 465 – 476. Leckie, D., Gougeon, F., Hill, D., Quinn, R., Armstrong, L., & Shreenan, R.

462

X. Yu et al. / Remote Sensing of Environment 90 (2004) 451–462

(2003). Combined high-density lidar and multispectral imagery for individual tree crown analysis. Canadian Journal of Remote Sensing, 29, 633 – 649. Lim, K., Treitz, P., Groot, A., & St-Onge, B. (2001). Estimation of individual tree heights using LIDAR remote sensing. Proceedings of the 23rd Annual Canadian Symposium on Remote Sensing, Quebec, QC, August 20 – 24, 2001. Maltamo, M., Mustonen, K., Hyyppa¨, J., Pitka¨nen, J., & Yu, X. (Submitted). The accuracy of estimating individual tree variables with airborne laser scanning in a boreal nature reserve. The Canadian Journal of Forest Research. Næsset, E., & Økland, T. (2002). Estimating tree height and tree crown properties using airborne scanning laser in a boreal nature reserve. Remote Sensing of Environment, 79, 105 – 115. Nelson, R., Krabill, W., & Tonelli, J. (1988). Estimating forest biomass and volume using airborne laser data. Remote Sensing of Environment, 24, 247 – 267. Nelson, R. F., Krabill, W. B., & Maclean, G. A. (1984). Determining forest canopy characteristics using airborne laser data. Remote Sensing of Environment, 15, 201 – 212. Nilsson, M. (1990). Forest inventory using an airborne LIDAR system. Proceedings of SNS/IUFRO workshop, Umea˚ , 26 – 28 February, 1990Report of Remote Sensing Laboratory ( pp. 133 – 139). Swedish University of Agricultural Sciences. Olsson, H. (1994). Changes in satellite-measured reflectances caused by thinning cutting in boreal forest. Remote Sensing of Environment, 50, 221 – 230. Pa¨ivinen, R., Nousiainen, M., & Korhonen, K. (1992). Puutunnusten mittaamisen luotettavuus (in Finnish). English summary: Accuracy of certain tree measurements. Folia Forestalia, 787 (18 pp.). Pekkarinen, A. (2002). Image segment-based spectral features in the estimation of timber volume. Remote Sensing of Environment, 82, 349 – 359. ˚ . (2001). Extraction of individual trees using laser radar data. Persson, A Go¨teborg, Sweden: Department of Signals and Systems, Chalmers University of Technology (28 pp.). ˚ ., Holmgren, J., & So¨derman, U. (2002). Detecting and measurPersson, A ing individual trees using an airborne laser scanner. Photogrammetric Engineering and Remote Sensing, 68, 925 – 932. Popescu, S. C., Wynne, R. H., & Nelson, R. E. (2002). Estimating plot-level tree heights with lidar: Local filtering with a canopy-height based variable window. Computers and Electronics in Agriculture, vol. 37(1 – 3) (pp. 71 – 95). Elsevier.

Pouliot, D. A., King, D. J., Bell, F. W., & Pitt, D. G. (2002). Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest regeneration. Remote Sensing of Environment, 82, 105 – 115. Quackenbush, L. J., Hopkins, P. F., & Kinn, G. J. (2000). Using template correlation to identify individual trees in high resolution imagery. Proceedings of ASPRS 2000 Annual Conference, Washington, DC. (CDROM). Ruppert, G., Wimmer, A., Beichel, R., & Ziegler, M. (2000). An adaptive multi-resolutional algorithm for high precision forest floor DTM generation. Proceedings of SPIE, laser radar technology and applications V, 26 – 28 April, Orlando, USA, vol. 4035 (12 pp.). Sironen, S., Kangas, A., Maltamo, M., & Kangas, J. (2003). Estimating individual tree growth with non-parametric methods. Canadian Journal of Forest Research, 33, 444 – 449. Stiteler, W. M. IV, & Hopkins, P. F. (2000). Using genetic algorithms to select tree crown templates for finding trees in digital imagery. Proceedings of ASPRS 2000 annual conference, Washington, DC. (CD-ROM). St-Onge, B. A. (2000). Estimating individual tree heights of the boreal forest using airborne laser altimetry and digital videography. Workshop of ISPRS WG III/2 & III/5: Mapping surface structure and topography by airborne and spaceborne lasers, 7 – 9.11.1999 La Jolla (California). International Archives of Photogrammetry and Remote Sensing, 32, 179 – 184. Tokola, T., & Heikkila¨, J. (1997). A priori site quality information in satellite image based forest inventory. Silva Fennica, 31, 67 – 78. Tomppo, E. (1991). Satellite image-based national forest inventory of Finland. International Archives of Photogrammetry and Remote Sensing, 28, 419 – 424. Tomppo, E. (1997). Recent status and further development of Finnish multisource forest inventory. The Marcus Wallenberg Foundation Symposia Proceedings, 11, 53 – 70. TopoSys, E. (1996). Digital elevation models, services and products. Rovensberg, Germany: TopoSys GmbH (10 pp.). Varjo, J. (1996). Controlling continuously updated forest data by satellite remote sensing. International Journal of Remote Sensing, 17, 43 – 67. Varjo, J., & Mery, G. (2001). Forest change detection. In M. Palo, J. Uusivuori, & G. Mery (Eds.), World Forests, Markets and Policies, vol. 3 (pp. 239 – 240). Kluwer Academic Publishing. Vehmas, M. (2002). Field data for laser scanner research (In Finnish). Report. University of Joensuu (9 pp.).