Directional Local Filtering for Stand Density Estimation ... - IEEE Xplore
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 10, NO. 4, JULY 2013
913
Directional Local Filtering for Stand Density Estimation in Closed Forest Canopies Using VHR Optical and LiDAR Data Frieke M. B. Van Coillie, Flore R. Devriendt, Lieven P. C. Verbeke, and Robert R. De Wulf
Abstract—In this letter, we present a novel object-based approach addressing individual tree crown (ITC) detection to assess stand density from remotely sensed imagery in closed forest canopies: directional local filtering (DLF). DLF is a variant of local maximum filtering (LMF). Within locally homogeneous areas, it uses a 1-D neighborhood and simultaneously searches for local directional maxima and minima. From the extracted local maxima and minima, a proxy for crown dimensions is inferred, which is in turn related to stand density. Developed on artificial imagery, the new object-based ITC method was tested on three different forest types in Belgium, which were all characterized by dense closed canopies: 1) a coniferous forest; 2) a mixed forest; and 3) a deciduous forest. Very high resolution aerial photographs, IKONOS imagery, and Light Detection and Ranging data, in conjunction with manually digitized and field survey data, were used to evaluate the new technique. The directional DLF approach yielded consistently stronger relations (in terms of R2 ) when compared with the conventional omnidirectional LMF technique. The qualitative evaluation clearly demonstrated that, next to stand density estimation, DLF also offered opportunities for full crown delineation. Index Terms—Closed forest canopies, filtering algorithms, high spatial resolution imaging, stand density.
I. I NTRODUCTION
S
TAND density, which is expressed as the number of trees per unit area, is an important forest management attribute. Many authors have used local spectral and textural image properties to estimate stand density (e.g., [1] and [2]). With the increasing availability of high spatial resolution data, remote sensing research in tree counting has focused directly on the key structural element of such images: the individual tree crown (ITC). ITC-based approaches can be divided into three main streams based on the primary type of information obtained: 1) tree location; 2) tree location and crown dimensions; or 3) full crown delineation [3]. Techniques for finding tree locaManuscript received September 14, 2012; revised December 20, 2012; accepted January 14, 2013. Date of publication February 12, 2013; date of current version May 27, 2013. This work was supported by the Belgian Science Policy Office. F. M. B. Van Coillie, F. R. Devriendt, and R. R. De Wulf are with the Laboratory of Forest Management and Spatial Information Techniques (FORSIT), Faculty of Bioscience Engineering, Ghent University, 9000 Ghent, Belgium (e-mail: [email protected]). L. P. C. Verbeke was with the Laboratory of Forest Management and Spatial Information Techniques (FORSIT), Ghent University, 9000 Ghent, Belgium. He is now with the Department of Information Technology (INTEC), Ghent University, 9050 Ledeberg, Belgium (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2013.2242044
tions are mostly based on detecting local maxima [4], [5]. Such techniques work rather well in medium to dense coniferous stands, provided that the detection filter kernel size is appropriate for the tree sizes and the image resolution [3]. In contrast to fixed window sizes [6] or to detect trees in more open areas, some authors developed local maxima filtering techniques with variable window sizes [7], [8]. Such techniques naturally lead to stem counts and are, thus, relatively reliable stand density estimators. However, the detection and count of deciduous trees is generally not as successful, as they may feature more than a single point of high brightness [3]. Techniques for detecting tree locations and crown dimensions are often based on retrieving local maxima and finding some edges of the crown [9]. Crown edges can be found using local transect analysis [10] or applying scale space theory [11]. Other more computer-intensive approaches to find tree locations and crown dimensions are based on template matching [12], Markov random fields [13], or marked point processes [14]. Techniques aiming at full crown delineation are either based on following valleys of shade between tree crowns [15]; following crown edges, as detected by a gradient operator, while analyzing their curvature [16]; 3-D modeling [17]; watershed segmentation [18]; or starting a region-growing segmentation from a seed point within a crown [19], [20]. In addition to tree detection and tree counting, because of their delineation of full crowns, these techniques are more appropriate for ITC-based species classifications, canopy closure estimations, canopy gap detection and distribution, and probably also for biomass or volume estimations [3]. Most of the aforementioned techniques were developed and assessed using high spatial resolution optical data, and are based on the assumption that there are peaks of reflectance around the tree tops and valleys along the canopy edges. However, the peaks and valleys are not always distinct since canopy reflectance is affected by various factors such as illumination conditions, canopy spectral properties, and complex canopy structure [21]. Recently, Light Detection and Ranging (LiDAR) data have emerged as important information source for ITC isolation and canopy information extraction [22], [23]. Compared with passive optical imaging, LiDAR has the advantage of directly measuring the 3-D coordinates of canopies. Therefore, the geometric, rather than spectral, peaks and valleys can be detected [21]. Hence, LiDAR-derived gridded canopy height models
1545-598X/$31.00 © 2013 IEEE
914
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 10, NO. 4, JULY 2013
(CHMs) are often processed using algorithms closely related to ITC methods designed for passive optical images [24]–[26]. A major shortcoming of many automatic ITC-based approaches is that, regardless of the algorithms or image types used, their accuracy is largely dependent upon the structural complexity and density of forest stands: high detection accuracy is attained in open single-story stands, whereas comparatively lower accuracy is attained in closed multistory stands [3], [27]. Falkowski et al. [28] attributed this to the fact that trees growing closely together often have overlapping or interlocking crowns, making it difficult to isolate individual trees. Furthermore, as forest canopy cover increases, trees in subdominant (i.e., intermediate or suppressed) canopy layers become occluded by the overstory, leading to low detection rates in these canopy layers [18], [21], [22]. From literature, it is apparent that, with the existing ITC-based methods, relatively good results can be achieved within managed forests, including natural or plantation conifer forests and orchards, with small species diversity and tree crowns that are typically symmetrical and circular in shape. However, in mixed-species forests of complex structure, where trees occur in multiple strata and are closely packed, detection and delineation accuracy often decreases drastically [20]. Stand density estimation based on ITC approaches, thus, remains a challenging research topic within unmanaged forest with closed canopies. The current study focuses on dense closed forest canopies and presents a new object-based ITC technique that belongs to the second category of finding tree locations and crown dimensions. Within locally homogeneous areas, directional local filtering (DLF) aims at detecting both local maxima and minima within a 1-D filter window. From the extracted local directional maxima and minima, a proxy for crown dimensions is inferred, which is in turn, object-wise, related to stand density. Stand density is no longer based on explicit individual tree identification but is based on a small-area statistic characterizing local density. Additionally, DLF offers opportunities for full crown delineation by, e.g., seeding a region-growing segmentation algorithm with the local maxima and bounding it by the local minima.
II. DLF A. DLF Motivation There are a number of different conceptual approaches to the problem of ITC detection, of which local maximum filtering (LMF) is by far the most frequently reported method. We present DLF as a variant of LMF and evaluate its performance to assess stand density. With DLF, we aim to refine LMF by using a 1-D window (line) and simultaneously searching for local directional minima. Both minima and maxima carry a physical meaning, namely, crown edges (valleys) and crown apexes (peaks), respectively. Filtering line-wise instead of windowwise offers the opportunity of finding an increasing amount of local directional extrema (see Fig. 1). It may be clear that the extracted local directional minima and maxima cannot be directly used to count tree stems. As further explained, a statistic of the distance between all two
Fig. 1.
LMF versus DLF: 2-D window versus 1-D line approach.
Fig. 2. Artificially generated tree lanes under different illumination conditions and (part) of the corresponding cross profiles (taken after conversion to gray values).
consecutive minima embracing a single maximum will be used as an object-based proxy for stand density. Working with different viewing angles was steered by experiments we performed on artificially generated tree lanes. Fig. 2 shows two identical artificial tree lanes that were rendered under perpendicular illumination conditions and cross profiles respectively parallel with and perpendicular to the illumination angle. From the profiles, it can be concluded that observation parallel with the illumination azimuth allows for individual tree identification. Looking perpendicular to the illumination azimuth results in a cross profile with reduced brightness dynamics, inducing the loss of some tree crowns. The more or less regular pattern of brightness peaks and valleys is no longer present in the perpendicular profile. Hence, one might benefit from limiting image analysis to directions parallel with the illumination vector, because this direction will yield the highest contrast possible. On the other hand, when spatial and radiometric resolutions are sufficiently high or with higher sun elevations, one might actually benefit from including both viewing directions in the analysis, since these two viewing directions then produce profiles with a large amount of uncorrelated information. These two extreme situations emphasize the importance of viewing direction when filtering line-wise. When applying DLF to LiDAR-derived CHMs, illumination
VAN COILLIE et al.: DIRECTIONAL LOCAL FILTERING FOR STAND DENSITY IN CLOSED CANOPIES
conditions are no longer an issue. However, filtering the mountainous CHM surface in two perpendicular directions does guarantee that all possible peaks and valleys are sampled. Likewise to LMF, the ability to identify individual trees with this technique is dependent upon careful selection of the 1-D filter length. The DLF filter length, thus, should be appropriate for the tree crown sizes and the images’ spatial resolution. B. DLF Implementation DLF was implemented in C++ using the Borland C++Builder (version 6.0) and operates on Microsoft Windows platforms. C. DLF Experimental Setup In a first stage, the DLF algorithm was developed using artificially generated very high resolution (VHR) optical images. Artificial image generation consisted of 1) modeling tree shapes, 2) generating multiple trees and assembling them into stands, and 3) illuminating these forests using the Persistence of Vision Raytracer rendering algorithm to produce simulated VHR optical imagery [2]. Using eight simulated images [size 1 ha, 1-m ground-sampled distance (GSD)] with varying number of trees (between 109 and 1176), DLF was tested on its capacity to estimate stand density. Crown dimensions were characterized by the distance between all two consecutive minima embracing a single maximum. A statistic of these distances was used as explanatory variable for stand density. In a second stage, DLF was applied to real VHR imagery over three different forest types in Flanders, the northern region of Belgium, which were all characterized by dense closed canopies: 1) a coniferous forest; 2) a mixed forest; and 3) a deciduous forest. The coniferous forest is part of a larger forest (650 ha) in Oud-Heverlee and is covered by a scanned color infrared aerial photograph (scale 1:5000, 1-m GSD). Tree tops were manually digitized in 20 selected forest stands spanning a range of densities. In total, 3161 tree tops were allocated as reference data, all of which are coniferous trees. The mixed forest, which is situated in the eastern province of Limburg, is covered by a panchromatic 1-m resolution IKONOS image mosaic (of three images) that is part of the global VHR coverage of the entire Flemish territory, which was recorded in the course of summer 2002. Reference data on trees species and density were provided by the Flemish Forest Inventory database for a systematic grid of circular plots with a radius of 18 m. Out of a total of 16 homogeneous samples plots, seven plots contained broadleaf species, eight plots consisted exclusively of conifers, and one plot contained a mix of coniferous and broadleaf species. The third forest site is the oldest forest reserve in Flanders, namely, Kersselaerspleyn. Kersselaerspleyn is part of the forest complex Sonian Forest, which is located in the south of Brussels and is dominated by beech and oak trees. Over this deciduous forest, LiDAR (Riegl LMS Q560 full-waveform laser scanner with point density > 10 points/m2 ) data were acquired during the summer of 2010. Subsequently, the LiDARderived CHM (0.5-m GSD, maximum return selection) was smoothed using a 3 × 3 median filter. Reference data on species composition, tree diameter, and stem position were measured
915
by the Flemish Research Institute for Nature and Forest (INBO) in the context of the monitoring program of the Flemish Forest Reserves [29]. From the systematic grid of circular plots with a radius of 30 m, 22 homogeneous plots were selected. Although the proposed method is presented on exemplar homogeneous forest plots, large-scale implementations would address locally homogeneous areas, i.e., image objects. III. R ESULTS AND D ISCUSSION A. Stage 1: Developing Stage In uniform fully stocked stands, it can be shown that stand density is a linear function of the inverse quadratic mean crown diameter. Based on the eight artificially rendered images, this linear relationship was validated. We related stand density to the inverted squared mean distance between all consecutive local minima embracing a single local maximum and found an R2 value of 0.87. Therefore, we decided to use this statistic as an explanatory variable for stand density. B. Stage 2: Application Stage We applied DLF to the aforementioned Flemish forest sites and compared DLF with LMF on their capacity to assess stand density. Therefore, we investigated the relation between true stand density and the inverted squared mean distance collected parallel with and perpendicular to the sun’s azimuth. Additionally, the average of the statistics in both directions was investigated. Finally, the number of local maxima per hectare detected by LMF were linearly related with true stand density. Whenever stand density was derived from an inventory database, trees in subdominant canopy layers were excluded as DLF and LMF can only identify trees from the upper canopy. 1) Coniferous Forest Site (Aerial Photograph): DLF filter length and LMF window size were set in a trial-and-error process to seven pixels and 7 × 7, respectively. R2 s for parallel and perpendicular directions are 0.80 and 0.63, respectively, demonstrating the higher information content of parallel directions. DLF generally outperformed LMF (R2 = 0.71), particularly when analyzing directions parallel with and perpendicular to the illumination vector were combined (R2 = 0.85) (see Fig. 3). 2) Mixed Forest Site (IKONOS Image): Based on the field reference database of the mixed forest site, a DLF filter length of seven pixels and an LMF window size of 7 × 7 were found appropriate. Roughly the same conclusions as aforementioned hold: DLF outperformed LMF (R2 = 0.58), and again, the parallel direction (R2 = 0.78) contained the most relevant information. Combined directions (R2 = 0.69), however, did not yield better correspondence compared with parallel directions (see Fig. 3). 3) Deciduous Forest Site (LiDAR Data): Based on ground survey reference information, DLF filter lengths and LMF window sizes were set for each individual circle plot. Filter lengths and window sizes varied between 4 and 14 pixels and between 4 × 4 and 14 × 14, respectively. In comparison with LMF (R2 = 0.73), DLF once more displayed improved performances, with higher R2 values in both the parallel (R2 = 0.83)
916
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 10, NO. 4, JULY 2013
Fig. 3. Comparison of the relation between stand density and (from left to right) the parallel, perpendicular, and combined DLF statistics and the number of local maxima per hectare, for the three Flemish forest sites.
and perpendicular (R2 = 0.76) directions (see Fig. 3). The latter was expected as with LiDAR data; illumination conditions are not an issue.
C. Qualitative Evaluation In order to evaluate DLF, Fig. 4 shows the DLF maxima (red) and minima (blue) calculated in four directions (0◦ –45◦ –90◦ –135◦ ) for a LiDAR CHM subset of one sample plot in Kersselaerspleyn. For comparison, the LMF maxima (yellow) and reference stem positions (green) are displayed. Fig. 4 clearly shows the potential of DLF as a new objectbased ITC technique of finding tree locations, as well as characterizing crown dimensions. The DLF minima neatly indicate the valleys between neighboring trees, demonstrating the possibilities DLF has for crown delineation. Additionally, the large branches typical of deciduous trees are well characterized by the DLF maxima. The reference stem positions are located close to the intersections of the DLF maxima, indicating the capacity of DLF for finding tree locations. Compared to LMF, with DLF, instead of one local maximum, considerable more measurements per crown are performed. These local directional maxima and minima are used to calculate a statistic (mean distance), which is integrated over a certain area (plot). Implementing DLF in different directions yields even more local maxima and minima, suggesting a more precise estimation of the object-based statistic that relates to stand density.
Fig. 4. Subset of a sample plot in Kersselaerspleyn (deciduous forest), showing the DLF maxima (red) and minima (blue) in four directions (0◦ –45◦ –90◦ –135◦ ) and the LMF maxima (yellow) and reference stem positions (green).
IV. C ONCLUSION In dense coniferous, as well as in deciduous and mixed forest canopies, DLF offers promise for stand density estimation. In terms of R2 , DLF consistently yielded stronger relationships with true stand density compared with LMF. The results on
VAN COILLIE et al.: DIRECTIONAL LOCAL FILTERING FOR STAND DENSITY IN CLOSED CANOPIES
aerial photographs, IKONOS images, and LiDAR data indicated that DLF is not only applicable to different forest types; it is also transferable to different image types. The qualitative evaluation clearly demonstrated that, next to tree positioning, DLF also offers opportunities for full crown delineation. ACKNOWLEDGMENT The authors would like to thank the Forest and Nature Department of the Ministry of the Flemish Community for providing the IKONOS imagery, as well as the database with sample plots on which the Flemish Forest Inventory (Afdeling Bos en Groen, 2001) is based; the Remote Sensing Laboratories (University of Zurich, Switzerland) and Pete Bunting (University of Aberystwyth, U.K.) for preprocessing the LiDAR data; and the Research Institute for Forest and Nature for collecting and providing the field reference data over Kersselaerspleyn. R EFERENCES [1] D. Klobucar and R. Pernar, “Artificial neural networks in the estimation of stand density from cyclic aerial photographs,” Sumarski List, vol. 133, no. 3/4, pp. 145–155, 2009. [2] F. M. B. Van Coillie, L. P. C. Verbeke, and R. R. De Wulf, “Waveletbased texture measures for semicontinuous stand density estimation from very high resolution optical imagery,” J. Appl. Remote Sens., vol. 5, no. 1, pp. 053560-1–053560-14, Jan. 2011. [3] F. A. Gougeon and D. G. Leckie, “The individual tree crown approach applied to Ikonos images of a coniferous plantation area,” Photogramm. Eng. Remote Sens., vol. 72, no. 11, pp. 1287–1297, 2006. [4] N. R. Eldridge and G. Edwards, “Acquiring localized forest inventory information: Extraction from high resolution airborne digital images,” in Proc. 16th Can. Symp. Remote Sens., Sherbrooke, QC, Canada, 1993, pp. 443–448. [5] K. Dralle and M. Rudemo, “Automatic estimation of individual tree positions from aerial photos,” Can. J. Forest Res.–Rev. Can. De Recherche Forestiere, vol. 27, no. 11, pp. 1728–1736, Nov. 1997. [6] P. Bolduc, K. Lowell, and G. Edwards, “Automated estimation of localized forest volume from large-scale aerial photographs and ancillary cartographic information in a boreal forest,” Int. J. Remote Sens., vol. 20, no. 18, pp. 3611–3624, Dec. 1999. [7] F. A. Gougeon and D. G. Leckie, “Forest regeneration: Individual tree crown detection techniques for density and stocking assessments,” in Proc. Autom. Interpret. High Spatial Resol. Digit. Imag. Forest., Int. Forum, D. A. Hill and D. G. Leckie, Eds., 1999, pp. 169–177. [8] M. Wulder, K. O. Niemann, and D. G. Goodenough, “Local maximum filtering for the extraction of tree locations and basal area from high spatial resolution imagery,” Remote Sens. Environ., vol. 73, no. 1, pp. 103–114, Jul. 2000. [9] J. Uuttera, A. Haara, T. Tokola, and M. Maltamo, “Determination of the spatial distribution of trees from digital aerial photographs,” Forest Ecol. Manage., vol. 110, no. 1–3, pp. 275–282, Oct. 1998. [10] D. A. Pouliot, D. J. King, F. W. Bell, and D. G. Pitt, “Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest regeneration,” Remote Sens. Environ., vol. 82, no. 2/3, pp. 322–334, Oct. 2002. [11] T. Brandtberg and F. Walter, “Automated delineation of individual tree crowns in high spatial resolution aerial images by multiple-scale analysis,” Mach. Vis. Appl., vol. 11, no. 2, pp. 64–73, Oct. 1998.
917
[12] R. Pollock, “The automatic recognition of individual trees in aerial images of forests based on a synthetic tree crown model,” Ph.D. dissertation, Dept. Comput. Sci., Univ. Brit. Columbia, Vancouver, BC, Canada, 1996. [13] X. Descombes and E. Pechersky, “Tree crown extraction using a three states Markov random field,” INRIA, Sophia Antipolis, France, Res. Rep. 5982, 2006. [14] G. Perrin, X. Descombes, and J. Zerubia, “2D and 3D vegetation resource parameters assessment using marked point processes,” in Proc. 18th Int. Conf. Pattern Recog., Y. Y. Tang, S. P. Wang, G. Lorette, D. S. Yeung, and H. Yan, Eds., 2006, vol. 1, pp. 1–4. [15] D. S. Culvenor, “TIDA: An algorithm for the delineation of tree crowns in high spatial resolution remotely sensed imagery,” Comput. Geosci., vol. 28, no. 1, pp. 33–44, Feb. 2002. [16] T. Brandtberg and F. Walter, “An algorithm for delineation of individual tree crowns in high spatial resolution aerial images using curved edge segments at multiple scales,” in Proc. Int. Forum Autom. Interpret. High Spatial Resol. Digit. Imag. Forest., D. A. Hill and D. G. Leckie, Eds., 1999, pp. 41–54. [17] P. Gong, Y. Sheng, and G. S. Biging, “3D model-based tree measurement from high-resolution aerial imagery,” Photogramm. Eng. Remote Sens., vol. 68, no. 11, pp. 1203–1212, Nov. 2002. [18] L. Wang, P. Gong, and G. S. Biging, “Individual tree-crown delineation and treetop detection high-spatial-resolution aerial imagery,” Photogramm. Eng. Remote Sens., vol. 70, no. 3, pp. 351–357, Mar. 2004. [19] M. Erikson and K. Olofsson, “Comparison of three individual tree crown detection methods,” Mach. Vis. Appl., vol. 16, no. 4, pp. 258–265, Sep. 2005. [20] P. Bunting and R. Lucas, “The delineation of tree crowns in Australian mixed species forests using hyperspectral Compact Airborne Spectrographic Imager (CASI) data,” Remote Sens. Environ., vol. 101, no. 2, pp. 230–248, Mar. 2006. [21] Q. Chen, D. Baldocchi, P. Gong, and M. Kelly, “Isolating individual trees in a savanna woodland using small footprint lidar data,” Photogramm. Eng. Remote Sens., vol. 72, no. 8, pp. 923–932, Aug. 2006. [22] A. Persson, J. Holmgren, and U. Soderman, “Detecting and measuring individual trees using an airborne laser scanner,” Photogramm. Eng. Remote Sens., vol. 68, no. 9, pp. 925–932, Sep. 2002. [23] S. C. Popescu, R. H. Wynne, and J. A. Scrivani, “Fusion of small-footprint lidar and multispectral data to estimate plot-level volume and biomass in deciduous and pine forests in Virginia, USA,” Forest Sci., vol. 50, no. 4, pp. 551–565, Aug. 2004. [24] S. C. Popescu and R. H. Wynne, “Seeing the trees in the forest: Using LiDAR and multispectral data fusion with local filtering and variable window size for estimating tree height,” Photogramm. Eng. Remote Sens., vol. 70, no. 5, pp. 589–604, May 2004. [25] B. Koch, U. Heyder, and H. Weinacker, “Detection of individual tree crowns in airborne lidar data,” Photogramm. Eng. Remote Sens., vol. 72, no. 4, pp. 357–363, Apr. 2006. [26] B. M. Wolf and C. Heipke, “Automatic extraction and delineation of single trees from remote sensing data,” Mach. Vis. Appl., vol. 18, no. 5, pp. 317– 330, Oct. 2007. [27] M. Maltamo, K. Mustonen, J. Hyyppa, J. Pitkanen, and X. Yu, “The accuracy of estimating individual tree variables with airborne laser scanning in a boreal nature reserve,” Can. J. Forest Res.–Rev. Can. De Recherche Forestiere, vol. 34, no. 9, pp. 1791–1801, Sep. 2004. [28] M. J. Falkowski, A. M. S. Smith, P. E. Gessler, A. T. Hudak, L. A. Vierling, and J. S. Evans, “The influence of conifer forest canopy cover on the accuracy of two individual tree measurement algorithms using lidar data,” Can. J. Remote Sens., vol. 34, no. 2, pp. S338–S350, 2008. [29] L. De Keersmaeker, H. Baeté, P. Van de Kerckhove, B. Bart Christiaens, M. Esprit, and K. Vandekerkhove, “Bosreservaat Kersselaerspleyn (Zoniënwoud)—Monitoringrapport: Monitoring van de vegetatie en de dendrometrische gegevens in de kernvlakte en de steekproefcirkels,” INBO, Eindhoven, The Netherlands, 2006.
913
Directional Local Filtering for Stand Density Estimation in Closed Forest Canopies Using VHR Optical and LiDAR Data Frieke M. B. Van Coillie, Flore R. Devriendt, Lieven P. C. Verbeke, and Robert R. De Wulf
Abstract—In this letter, we present a novel object-based approach addressing individual tree crown (ITC) detection to assess stand density from remotely sensed imagery in closed forest canopies: directional local filtering (DLF). DLF is a variant of local maximum filtering (LMF). Within locally homogeneous areas, it uses a 1-D neighborhood and simultaneously searches for local directional maxima and minima. From the extracted local maxima and minima, a proxy for crown dimensions is inferred, which is in turn related to stand density. Developed on artificial imagery, the new object-based ITC method was tested on three different forest types in Belgium, which were all characterized by dense closed canopies: 1) a coniferous forest; 2) a mixed forest; and 3) a deciduous forest. Very high resolution aerial photographs, IKONOS imagery, and Light Detection and Ranging data, in conjunction with manually digitized and field survey data, were used to evaluate the new technique. The directional DLF approach yielded consistently stronger relations (in terms of R2 ) when compared with the conventional omnidirectional LMF technique. The qualitative evaluation clearly demonstrated that, next to stand density estimation, DLF also offered opportunities for full crown delineation. Index Terms—Closed forest canopies, filtering algorithms, high spatial resolution imaging, stand density.
I. I NTRODUCTION
S
TAND density, which is expressed as the number of trees per unit area, is an important forest management attribute. Many authors have used local spectral and textural image properties to estimate stand density (e.g., [1] and [2]). With the increasing availability of high spatial resolution data, remote sensing research in tree counting has focused directly on the key structural element of such images: the individual tree crown (ITC). ITC-based approaches can be divided into three main streams based on the primary type of information obtained: 1) tree location; 2) tree location and crown dimensions; or 3) full crown delineation [3]. Techniques for finding tree locaManuscript received September 14, 2012; revised December 20, 2012; accepted January 14, 2013. Date of publication February 12, 2013; date of current version May 27, 2013. This work was supported by the Belgian Science Policy Office. F. M. B. Van Coillie, F. R. Devriendt, and R. R. De Wulf are with the Laboratory of Forest Management and Spatial Information Techniques (FORSIT), Faculty of Bioscience Engineering, Ghent University, 9000 Ghent, Belgium (e-mail: [email protected]). L. P. C. Verbeke was with the Laboratory of Forest Management and Spatial Information Techniques (FORSIT), Ghent University, 9000 Ghent, Belgium. He is now with the Department of Information Technology (INTEC), Ghent University, 9050 Ledeberg, Belgium (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2013.2242044
tions are mostly based on detecting local maxima [4], [5]. Such techniques work rather well in medium to dense coniferous stands, provided that the detection filter kernel size is appropriate for the tree sizes and the image resolution [3]. In contrast to fixed window sizes [6] or to detect trees in more open areas, some authors developed local maxima filtering techniques with variable window sizes [7], [8]. Such techniques naturally lead to stem counts and are, thus, relatively reliable stand density estimators. However, the detection and count of deciduous trees is generally not as successful, as they may feature more than a single point of high brightness [3]. Techniques for detecting tree locations and crown dimensions are often based on retrieving local maxima and finding some edges of the crown [9]. Crown edges can be found using local transect analysis [10] or applying scale space theory [11]. Other more computer-intensive approaches to find tree locations and crown dimensions are based on template matching [12], Markov random fields [13], or marked point processes [14]. Techniques aiming at full crown delineation are either based on following valleys of shade between tree crowns [15]; following crown edges, as detected by a gradient operator, while analyzing their curvature [16]; 3-D modeling [17]; watershed segmentation [18]; or starting a region-growing segmentation from a seed point within a crown [19], [20]. In addition to tree detection and tree counting, because of their delineation of full crowns, these techniques are more appropriate for ITC-based species classifications, canopy closure estimations, canopy gap detection and distribution, and probably also for biomass or volume estimations [3]. Most of the aforementioned techniques were developed and assessed using high spatial resolution optical data, and are based on the assumption that there are peaks of reflectance around the tree tops and valleys along the canopy edges. However, the peaks and valleys are not always distinct since canopy reflectance is affected by various factors such as illumination conditions, canopy spectral properties, and complex canopy structure [21]. Recently, Light Detection and Ranging (LiDAR) data have emerged as important information source for ITC isolation and canopy information extraction [22], [23]. Compared with passive optical imaging, LiDAR has the advantage of directly measuring the 3-D coordinates of canopies. Therefore, the geometric, rather than spectral, peaks and valleys can be detected [21]. Hence, LiDAR-derived gridded canopy height models
1545-598X/$31.00 © 2013 IEEE
914
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 10, NO. 4, JULY 2013
(CHMs) are often processed using algorithms closely related to ITC methods designed for passive optical images [24]–[26]. A major shortcoming of many automatic ITC-based approaches is that, regardless of the algorithms or image types used, their accuracy is largely dependent upon the structural complexity and density of forest stands: high detection accuracy is attained in open single-story stands, whereas comparatively lower accuracy is attained in closed multistory stands [3], [27]. Falkowski et al. [28] attributed this to the fact that trees growing closely together often have overlapping or interlocking crowns, making it difficult to isolate individual trees. Furthermore, as forest canopy cover increases, trees in subdominant (i.e., intermediate or suppressed) canopy layers become occluded by the overstory, leading to low detection rates in these canopy layers [18], [21], [22]. From literature, it is apparent that, with the existing ITC-based methods, relatively good results can be achieved within managed forests, including natural or plantation conifer forests and orchards, with small species diversity and tree crowns that are typically symmetrical and circular in shape. However, in mixed-species forests of complex structure, where trees occur in multiple strata and are closely packed, detection and delineation accuracy often decreases drastically [20]. Stand density estimation based on ITC approaches, thus, remains a challenging research topic within unmanaged forest with closed canopies. The current study focuses on dense closed forest canopies and presents a new object-based ITC technique that belongs to the second category of finding tree locations and crown dimensions. Within locally homogeneous areas, directional local filtering (DLF) aims at detecting both local maxima and minima within a 1-D filter window. From the extracted local directional maxima and minima, a proxy for crown dimensions is inferred, which is in turn, object-wise, related to stand density. Stand density is no longer based on explicit individual tree identification but is based on a small-area statistic characterizing local density. Additionally, DLF offers opportunities for full crown delineation by, e.g., seeding a region-growing segmentation algorithm with the local maxima and bounding it by the local minima.
II. DLF A. DLF Motivation There are a number of different conceptual approaches to the problem of ITC detection, of which local maximum filtering (LMF) is by far the most frequently reported method. We present DLF as a variant of LMF and evaluate its performance to assess stand density. With DLF, we aim to refine LMF by using a 1-D window (line) and simultaneously searching for local directional minima. Both minima and maxima carry a physical meaning, namely, crown edges (valleys) and crown apexes (peaks), respectively. Filtering line-wise instead of windowwise offers the opportunity of finding an increasing amount of local directional extrema (see Fig. 1). It may be clear that the extracted local directional minima and maxima cannot be directly used to count tree stems. As further explained, a statistic of the distance between all two
Fig. 1.
LMF versus DLF: 2-D window versus 1-D line approach.
Fig. 2. Artificially generated tree lanes under different illumination conditions and (part) of the corresponding cross profiles (taken after conversion to gray values).
consecutive minima embracing a single maximum will be used as an object-based proxy for stand density. Working with different viewing angles was steered by experiments we performed on artificially generated tree lanes. Fig. 2 shows two identical artificial tree lanes that were rendered under perpendicular illumination conditions and cross profiles respectively parallel with and perpendicular to the illumination angle. From the profiles, it can be concluded that observation parallel with the illumination azimuth allows for individual tree identification. Looking perpendicular to the illumination azimuth results in a cross profile with reduced brightness dynamics, inducing the loss of some tree crowns. The more or less regular pattern of brightness peaks and valleys is no longer present in the perpendicular profile. Hence, one might benefit from limiting image analysis to directions parallel with the illumination vector, because this direction will yield the highest contrast possible. On the other hand, when spatial and radiometric resolutions are sufficiently high or with higher sun elevations, one might actually benefit from including both viewing directions in the analysis, since these two viewing directions then produce profiles with a large amount of uncorrelated information. These two extreme situations emphasize the importance of viewing direction when filtering line-wise. When applying DLF to LiDAR-derived CHMs, illumination
VAN COILLIE et al.: DIRECTIONAL LOCAL FILTERING FOR STAND DENSITY IN CLOSED CANOPIES
conditions are no longer an issue. However, filtering the mountainous CHM surface in two perpendicular directions does guarantee that all possible peaks and valleys are sampled. Likewise to LMF, the ability to identify individual trees with this technique is dependent upon careful selection of the 1-D filter length. The DLF filter length, thus, should be appropriate for the tree crown sizes and the images’ spatial resolution. B. DLF Implementation DLF was implemented in C++ using the Borland C++Builder (version 6.0) and operates on Microsoft Windows platforms. C. DLF Experimental Setup In a first stage, the DLF algorithm was developed using artificially generated very high resolution (VHR) optical images. Artificial image generation consisted of 1) modeling tree shapes, 2) generating multiple trees and assembling them into stands, and 3) illuminating these forests using the Persistence of Vision Raytracer rendering algorithm to produce simulated VHR optical imagery [2]. Using eight simulated images [size 1 ha, 1-m ground-sampled distance (GSD)] with varying number of trees (between 109 and 1176), DLF was tested on its capacity to estimate stand density. Crown dimensions were characterized by the distance between all two consecutive minima embracing a single maximum. A statistic of these distances was used as explanatory variable for stand density. In a second stage, DLF was applied to real VHR imagery over three different forest types in Flanders, the northern region of Belgium, which were all characterized by dense closed canopies: 1) a coniferous forest; 2) a mixed forest; and 3) a deciduous forest. The coniferous forest is part of a larger forest (650 ha) in Oud-Heverlee and is covered by a scanned color infrared aerial photograph (scale 1:5000, 1-m GSD). Tree tops were manually digitized in 20 selected forest stands spanning a range of densities. In total, 3161 tree tops were allocated as reference data, all of which are coniferous trees. The mixed forest, which is situated in the eastern province of Limburg, is covered by a panchromatic 1-m resolution IKONOS image mosaic (of three images) that is part of the global VHR coverage of the entire Flemish territory, which was recorded in the course of summer 2002. Reference data on trees species and density were provided by the Flemish Forest Inventory database for a systematic grid of circular plots with a radius of 18 m. Out of a total of 16 homogeneous samples plots, seven plots contained broadleaf species, eight plots consisted exclusively of conifers, and one plot contained a mix of coniferous and broadleaf species. The third forest site is the oldest forest reserve in Flanders, namely, Kersselaerspleyn. Kersselaerspleyn is part of the forest complex Sonian Forest, which is located in the south of Brussels and is dominated by beech and oak trees. Over this deciduous forest, LiDAR (Riegl LMS Q560 full-waveform laser scanner with point density > 10 points/m2 ) data were acquired during the summer of 2010. Subsequently, the LiDARderived CHM (0.5-m GSD, maximum return selection) was smoothed using a 3 × 3 median filter. Reference data on species composition, tree diameter, and stem position were measured
915
by the Flemish Research Institute for Nature and Forest (INBO) in the context of the monitoring program of the Flemish Forest Reserves [29]. From the systematic grid of circular plots with a radius of 30 m, 22 homogeneous plots were selected. Although the proposed method is presented on exemplar homogeneous forest plots, large-scale implementations would address locally homogeneous areas, i.e., image objects. III. R ESULTS AND D ISCUSSION A. Stage 1: Developing Stage In uniform fully stocked stands, it can be shown that stand density is a linear function of the inverse quadratic mean crown diameter. Based on the eight artificially rendered images, this linear relationship was validated. We related stand density to the inverted squared mean distance between all consecutive local minima embracing a single local maximum and found an R2 value of 0.87. Therefore, we decided to use this statistic as an explanatory variable for stand density. B. Stage 2: Application Stage We applied DLF to the aforementioned Flemish forest sites and compared DLF with LMF on their capacity to assess stand density. Therefore, we investigated the relation between true stand density and the inverted squared mean distance collected parallel with and perpendicular to the sun’s azimuth. Additionally, the average of the statistics in both directions was investigated. Finally, the number of local maxima per hectare detected by LMF were linearly related with true stand density. Whenever stand density was derived from an inventory database, trees in subdominant canopy layers were excluded as DLF and LMF can only identify trees from the upper canopy. 1) Coniferous Forest Site (Aerial Photograph): DLF filter length and LMF window size were set in a trial-and-error process to seven pixels and 7 × 7, respectively. R2 s for parallel and perpendicular directions are 0.80 and 0.63, respectively, demonstrating the higher information content of parallel directions. DLF generally outperformed LMF (R2 = 0.71), particularly when analyzing directions parallel with and perpendicular to the illumination vector were combined (R2 = 0.85) (see Fig. 3). 2) Mixed Forest Site (IKONOS Image): Based on the field reference database of the mixed forest site, a DLF filter length of seven pixels and an LMF window size of 7 × 7 were found appropriate. Roughly the same conclusions as aforementioned hold: DLF outperformed LMF (R2 = 0.58), and again, the parallel direction (R2 = 0.78) contained the most relevant information. Combined directions (R2 = 0.69), however, did not yield better correspondence compared with parallel directions (see Fig. 3). 3) Deciduous Forest Site (LiDAR Data): Based on ground survey reference information, DLF filter lengths and LMF window sizes were set for each individual circle plot. Filter lengths and window sizes varied between 4 and 14 pixels and between 4 × 4 and 14 × 14, respectively. In comparison with LMF (R2 = 0.73), DLF once more displayed improved performances, with higher R2 values in both the parallel (R2 = 0.83)
916
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 10, NO. 4, JULY 2013
Fig. 3. Comparison of the relation between stand density and (from left to right) the parallel, perpendicular, and combined DLF statistics and the number of local maxima per hectare, for the three Flemish forest sites.
and perpendicular (R2 = 0.76) directions (see Fig. 3). The latter was expected as with LiDAR data; illumination conditions are not an issue.
C. Qualitative Evaluation In order to evaluate DLF, Fig. 4 shows the DLF maxima (red) and minima (blue) calculated in four directions (0◦ –45◦ –90◦ –135◦ ) for a LiDAR CHM subset of one sample plot in Kersselaerspleyn. For comparison, the LMF maxima (yellow) and reference stem positions (green) are displayed. Fig. 4 clearly shows the potential of DLF as a new objectbased ITC technique of finding tree locations, as well as characterizing crown dimensions. The DLF minima neatly indicate the valleys between neighboring trees, demonstrating the possibilities DLF has for crown delineation. Additionally, the large branches typical of deciduous trees are well characterized by the DLF maxima. The reference stem positions are located close to the intersections of the DLF maxima, indicating the capacity of DLF for finding tree locations. Compared to LMF, with DLF, instead of one local maximum, considerable more measurements per crown are performed. These local directional maxima and minima are used to calculate a statistic (mean distance), which is integrated over a certain area (plot). Implementing DLF in different directions yields even more local maxima and minima, suggesting a more precise estimation of the object-based statistic that relates to stand density.
Fig. 4. Subset of a sample plot in Kersselaerspleyn (deciduous forest), showing the DLF maxima (red) and minima (blue) in four directions (0◦ –45◦ –90◦ –135◦ ) and the LMF maxima (yellow) and reference stem positions (green).
IV. C ONCLUSION In dense coniferous, as well as in deciduous and mixed forest canopies, DLF offers promise for stand density estimation. In terms of R2 , DLF consistently yielded stronger relationships with true stand density compared with LMF. The results on
VAN COILLIE et al.: DIRECTIONAL LOCAL FILTERING FOR STAND DENSITY IN CLOSED CANOPIES
aerial photographs, IKONOS images, and LiDAR data indicated that DLF is not only applicable to different forest types; it is also transferable to different image types. The qualitative evaluation clearly demonstrated that, next to tree positioning, DLF also offers opportunities for full crown delineation. ACKNOWLEDGMENT The authors would like to thank the Forest and Nature Department of the Ministry of the Flemish Community for providing the IKONOS imagery, as well as the database with sample plots on which the Flemish Forest Inventory (Afdeling Bos en Groen, 2001) is based; the Remote Sensing Laboratories (University of Zurich, Switzerland) and Pete Bunting (University of Aberystwyth, U.K.) for preprocessing the LiDAR data; and the Research Institute for Forest and Nature for collecting and providing the field reference data over Kersselaerspleyn. R EFERENCES [1] D. Klobucar and R. Pernar, “Artificial neural networks in the estimation of stand density from cyclic aerial photographs,” Sumarski List, vol. 133, no. 3/4, pp. 145–155, 2009. [2] F. M. B. Van Coillie, L. P. C. Verbeke, and R. R. De Wulf, “Waveletbased texture measures for semicontinuous stand density estimation from very high resolution optical imagery,” J. Appl. Remote Sens., vol. 5, no. 1, pp. 053560-1–053560-14, Jan. 2011. [3] F. A. Gougeon and D. G. Leckie, “The individual tree crown approach applied to Ikonos images of a coniferous plantation area,” Photogramm. Eng. Remote Sens., vol. 72, no. 11, pp. 1287–1297, 2006. [4] N. R. Eldridge and G. Edwards, “Acquiring localized forest inventory information: Extraction from high resolution airborne digital images,” in Proc. 16th Can. Symp. Remote Sens., Sherbrooke, QC, Canada, 1993, pp. 443–448. [5] K. Dralle and M. Rudemo, “Automatic estimation of individual tree positions from aerial photos,” Can. J. Forest Res.–Rev. Can. De Recherche Forestiere, vol. 27, no. 11, pp. 1728–1736, Nov. 1997. [6] P. Bolduc, K. Lowell, and G. Edwards, “Automated estimation of localized forest volume from large-scale aerial photographs and ancillary cartographic information in a boreal forest,” Int. J. Remote Sens., vol. 20, no. 18, pp. 3611–3624, Dec. 1999. [7] F. A. Gougeon and D. G. Leckie, “Forest regeneration: Individual tree crown detection techniques for density and stocking assessments,” in Proc. Autom. Interpret. High Spatial Resol. Digit. Imag. Forest., Int. Forum, D. A. Hill and D. G. Leckie, Eds., 1999, pp. 169–177. [8] M. Wulder, K. O. Niemann, and D. G. Goodenough, “Local maximum filtering for the extraction of tree locations and basal area from high spatial resolution imagery,” Remote Sens. Environ., vol. 73, no. 1, pp. 103–114, Jul. 2000. [9] J. Uuttera, A. Haara, T. Tokola, and M. Maltamo, “Determination of the spatial distribution of trees from digital aerial photographs,” Forest Ecol. Manage., vol. 110, no. 1–3, pp. 275–282, Oct. 1998. [10] D. A. Pouliot, D. J. King, F. W. Bell, and D. G. Pitt, “Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest regeneration,” Remote Sens. Environ., vol. 82, no. 2/3, pp. 322–334, Oct. 2002. [11] T. Brandtberg and F. Walter, “Automated delineation of individual tree crowns in high spatial resolution aerial images by multiple-scale analysis,” Mach. Vis. Appl., vol. 11, no. 2, pp. 64–73, Oct. 1998.
917
[12] R. Pollock, “The automatic recognition of individual trees in aerial images of forests based on a synthetic tree crown model,” Ph.D. dissertation, Dept. Comput. Sci., Univ. Brit. Columbia, Vancouver, BC, Canada, 1996. [13] X. Descombes and E. Pechersky, “Tree crown extraction using a three states Markov random field,” INRIA, Sophia Antipolis, France, Res. Rep. 5982, 2006. [14] G. Perrin, X. Descombes, and J. Zerubia, “2D and 3D vegetation resource parameters assessment using marked point processes,” in Proc. 18th Int. Conf. Pattern Recog., Y. Y. Tang, S. P. Wang, G. Lorette, D. S. Yeung, and H. Yan, Eds., 2006, vol. 1, pp. 1–4. [15] D. S. Culvenor, “TIDA: An algorithm for the delineation of tree crowns in high spatial resolution remotely sensed imagery,” Comput. Geosci., vol. 28, no. 1, pp. 33–44, Feb. 2002. [16] T. Brandtberg and F. Walter, “An algorithm for delineation of individual tree crowns in high spatial resolution aerial images using curved edge segments at multiple scales,” in Proc. Int. Forum Autom. Interpret. High Spatial Resol. Digit. Imag. Forest., D. A. Hill and D. G. Leckie, Eds., 1999, pp. 41–54. [17] P. Gong, Y. Sheng, and G. S. Biging, “3D model-based tree measurement from high-resolution aerial imagery,” Photogramm. Eng. Remote Sens., vol. 68, no. 11, pp. 1203–1212, Nov. 2002. [18] L. Wang, P. Gong, and G. S. Biging, “Individual tree-crown delineation and treetop detection high-spatial-resolution aerial imagery,” Photogramm. Eng. Remote Sens., vol. 70, no. 3, pp. 351–357, Mar. 2004. [19] M. Erikson and K. Olofsson, “Comparison of three individual tree crown detection methods,” Mach. Vis. Appl., vol. 16, no. 4, pp. 258–265, Sep. 2005. [20] P. Bunting and R. Lucas, “The delineation of tree crowns in Australian mixed species forests using hyperspectral Compact Airborne Spectrographic Imager (CASI) data,” Remote Sens. Environ., vol. 101, no. 2, pp. 230–248, Mar. 2006. [21] Q. Chen, D. Baldocchi, P. Gong, and M. Kelly, “Isolating individual trees in a savanna woodland using small footprint lidar data,” Photogramm. Eng. Remote Sens., vol. 72, no. 8, pp. 923–932, Aug. 2006. [22] A. Persson, J. Holmgren, and U. Soderman, “Detecting and measuring individual trees using an airborne laser scanner,” Photogramm. Eng. Remote Sens., vol. 68, no. 9, pp. 925–932, Sep. 2002. [23] S. C. Popescu, R. H. Wynne, and J. A. Scrivani, “Fusion of small-footprint lidar and multispectral data to estimate plot-level volume and biomass in deciduous and pine forests in Virginia, USA,” Forest Sci., vol. 50, no. 4, pp. 551–565, Aug. 2004. [24] S. C. Popescu and R. H. Wynne, “Seeing the trees in the forest: Using LiDAR and multispectral data fusion with local filtering and variable window size for estimating tree height,” Photogramm. Eng. Remote Sens., vol. 70, no. 5, pp. 589–604, May 2004. [25] B. Koch, U. Heyder, and H. Weinacker, “Detection of individual tree crowns in airborne lidar data,” Photogramm. Eng. Remote Sens., vol. 72, no. 4, pp. 357–363, Apr. 2006. [26] B. M. Wolf and C. Heipke, “Automatic extraction and delineation of single trees from remote sensing data,” Mach. Vis. Appl., vol. 18, no. 5, pp. 317– 330, Oct. 2007. [27] M. Maltamo, K. Mustonen, J. Hyyppa, J. Pitkanen, and X. Yu, “The accuracy of estimating individual tree variables with airborne laser scanning in a boreal nature reserve,” Can. J. Forest Res.–Rev. Can. De Recherche Forestiere, vol. 34, no. 9, pp. 1791–1801, Sep. 2004. [28] M. J. Falkowski, A. M. S. Smith, P. E. Gessler, A. T. Hudak, L. A. Vierling, and J. S. Evans, “The influence of conifer forest canopy cover on the accuracy of two individual tree measurement algorithms using lidar data,” Can. J. Remote Sens., vol. 34, no. 2, pp. S338–S350, 2008. [29] L. De Keersmaeker, H. Baeté, P. Van de Kerckhove, B. Bart Christiaens, M. Esprit, and K. Vandekerkhove, “Bosreservaat Kersselaerspleyn (Zoniënwoud)—Monitoringrapport: Monitoring van de vegetatie en de dendrometrische gegevens in de kernvlakte en de steekproefcirkels,” INBO, Eindhoven, The Netherlands, 2006.
