Detection with Crown Delineation and Classification on Homogeneity Guided Smoothed High Resolution (50 cm) Multispectral Airborne Digital Data

Abstract: A method of counting the number of coniferous trees by species within forest compartments was developed by combining an individual tree crown delineation technique with a treetop detection technique, using high spatial resolution optical sensor data. When this method was verified against field data from the Shinshu University Campus Forest composed of various cover types, the accuracy for the total number of trees per stand was higher than 84%. This shows improvements over the individual tree crown delineation technique alone which had accuracies lower than 62%, or the treetop detection technique alone which had accuracies lower than 78%. However, the accuracy of the number of trees classified by species was less than 84%. The total number of trees by species per stand was improved with exclusion of the understory species and ranged from 45.2% to 93.8% for Chamaecyparis obtusa and C. pisifera and from 37.9% to 98.1% for broad-leaved trees because many of these were understory species. The better overall results are attributable primarily to the overestimation of


Introduction
Forests in Japan cover approximately 25.10 million ha or 66% of the total land area.Of these, 10 million ha (~40%) are artificial and plantations over 50 years account for about 35% of this area.The majority of plantations are dominated by conifers, with Cryptomeria japonica (Japanese Cedar), Chamaecyparis obtuse (Hinoki cypress), Chamaecyparis pisifera (Sawara cypress), Pinus densiflora (Japanese red pine), and Larix kaempferi (Japanese larch) being commonplace.Plantations aged between 31 and 60 years are managed by thinning or selective cutting or are harvested for timber, biomass and provision of clean energy [1].Forest officers and landowners require accurate information on the number and distribution of tree species to support management of their plantation resources.The number of trees and distribution of each tree species are basic information that can be used to evaluate current stand conditions in forest management.
At present, field surveys for forest resource management count the number of trees and measure the diameter at breast height (DBH) in small sample plots (0.1 ha).Usually 3-5 plots are located in each compartment.The structure of the entire forest resource is estimated by multiplying these measured values by the total forest area.However, this method is less accurate in large forests where stand conditions, species and stem densities vary.Although the number of study plots could be increased for greater accuracy, this would require a great deal of labor and cost.In addition, management operations have been abandoned in some forests following timber price decreases and as landowners age and retire.Forestry officers and landowners therefore need more accurate information on forest resource conditions.
The purpose of this study was to develop an economical method for estimating the number of trees of each dominant species over large areas using remote sensing data.Satellites are advantageous, as they can continuously capture data in multiple areas.The commercial satellites IKONOS, QuickBird, Geoeye-1 and WorldView-2, which have high spatial resolutions of 1 m or less in panchromatic mode, were launched in 1999, 2001, 2008 and 2009 respectively.As a result, the acquisition of detailed forest information from space has become possible and may be a more effective way to assess large areas of forest cover.Beginning in 2009, airborne remote sensing with high-resolution digital images taken over five years throughout the whole forest area of Japan has replaced previous analog aerial photographs.
Forestry studies with high-resolution airborne data have been successful in Canada and the United States [2][3][4][5].Tree counting, tree-crown delineation, species identification, crown density estimation and forest stand polygon delineation have been made possible with high-resolution data such as that collected via the airborne Multi-detector Electro-optical Imaging Sensor (MEIS), the Compact Airborne Spectrographic Imager (CASI) and the Leica Airborne Digital Sensor (ADS)-40 or 80. Tree counting techniques are based mainly on detecting local maxima in a smoothed image of the coniferous forested area [6][7][8], provided that the detection filter size and smoothing of the image are appropriate to the tree size and image spatial resolution [9][10][11].
Techniques for tree-crown delineation are often based on first finding local maxima and then locating crown edges.A fundamental assumption inherent to crown delineation methods is that the main part of a crown is brighter than the lower edge of the crown, particularly at the boundary between crowns.Methods for delineating tree crowns include three main approaches: bottom-up, top-down, and template-matching algorithms.Top-down algorithms can be divided into watershed, multiple-scale edge segments, threshold-based spatial clustering, and double aspect methods.The template-matching algorithms match a synthetic image model or template of tree crowns with the image radiometric values [12][13][14].The valley-following method [15] used in the present study is a bottom-up algorithm.Many examples of its application can be found in mature Canadian conifer forests.
In Japan, tree counting and tree-crown delineation methods have been applied to high-resolution airborne data from Japanese conifer plantations [16].In this study, the accuracy of the image analysis was verified by comparing tree counts or maps of delineated crowns to detailed field-based inventories.This approach is more effective in delineating tree crowns forming the upper canopy of plantations comprised purely of coniferous species, although accuracies were <60%.
A representation of how tree crowns and their tops are viewed from the ground and from overhead is given in Figure 1.Here, three tree crowns seen from above are mistakenly counted as one large tree crown.On the other hand, tree numbers obtained by the tree counting method (treetop) often exceed the number of trees in field surveys, as spurious multiple treetops are often extracted from single tree-crowns or from canopy gaps showing the forest floor.Furthermore, the identification of treetops from point data is less reliable.
Thus, we devised an improved method of extracting the number of trees of each species by combining the advantages of both methods.Furthermore, this method of automatically counting trees by species within forest polygons was developed using the framework of commonly available GIS software, because the digitization of forest information has progressed with the use of GIS in Japan.-21 2.
The original images with sensor orientation, topographic relief displacement, and systematic errors were rectified by ortho-correction based on X, Y, and Z values for sensor positions derived from a GPS system and the principal points and focal length of the sensor.The geographic projection was converted to Japan JGD2000 Zone 8.The ground resolution of the multi-image was 50 cm after correction.Image processing and analysis were performed using ortho-correction processing with ERDAS IMAGINE 8.6 [17] and a tree-based stand map on field data, which was processed with ArcGIS 9.3 [18].Treetop detection and tree-crown delineations were performed using the Individual Tree Crown (ITC)-Suite [19] and Geomatica 9 [20].

Field Data Collection
The field investigation was conducted from April 2005 to June 2010.Detailed stand information was available for forest compartments 1-7, where all trees ≥8 cm in diameter at breast height (DBH) were measured.The survey noted species, DBH, height (H), tree position, and strata (upper, intermediate, or understory layer).The two-story plantations, in which young C. obtusa or C. japonica were planted under older P. densiflora stands, were high density.Species, age, compartment area, count, average DBH and height, timber volume, and growth per hectare were summarized by forest compartment (Table 2).C. obtusa, P. densiflora, and L. kaempferi stands ranged in age from 24 to 86 years; in timber volume from 426 to 509 m 3 ; and in crown density from seven to nine, with high timber volume and high tree density.

GIS Data
Forest compartment boundaries, forest roads, forest survey data, and geographic data such as contour lines on the forest base map (scale 1:5,000) of the campus research forest were included in the forest database.All tree positions in forest compartments 1-7 from the field survey were transferred and registered in the forest database using GIS [16].

Methods
The flowchart in Figure 3 provides an overview of methods.The general steps used in this study were as follows: Figure 3. Overview of the approach.
Forested areas were extracted using forest boundaries from the GIS database to separate them from non-forested areas, such as roads, grass areas, and artificial structures.
(1) Homogeneous cover-type regions were separated by textures or gray levels, which were reflecting tree species and crown sizes.The diameters of the tree crowns ranged from 3 to 10 m; small tree areas were designated as those with tree crowns of 3 to 5 m, and large tree areas were those with crowns of 5 or more meters.Next, small tree-crown areas were smoothed on the image with a 3 × 3 average pixel filter; and large tree areas with an additional 5 × 5 pixel filter.(4) ITCs were classified into species using a supervised classification process based on comparing multi-spectral crown signatures with a maximum likelihood decision rule [16,19].(5) The existing two methods of "Tree top detection" and "ITCs species classification" were combined in the new technique.Tree top pixels were identified to species by overlaying them onto the ITC classification and thus, spurious treetops from canopy gaps or shaded areas were removed.These pixels were transferred from a pixel to point data using "raster to vector" conversion function, and then automatically labeled.(6) The number of trees by species within forest stand polygons was semi-automatically counted.(7) The data derived from these techniques were registered in a forest database and compared with the field investigation data to verify the effectiveness of this approach.

Homogeneity Cover Type
Homogeneous cover type analysis separated regions with different textures or gray levels by tree species and crown size.Small tree areas were extracted by blocking out large tree areas (Figure 4).Small tree crown areas were smoothed separately with a 3 × 3 average filter, and large tree areas were smoothed with a 3 × 3 and then a 5 × 5 averaging filter.

Identification of Tree Species
The ITC isolation image was produced using valley-following procedures.The ITC isolation image and multi-spectral images were fed into a supervised classification process to identify species [16].Treetops were overlain onto the ITC classification image (Figure 5(a)) and then associated with a species semi-automatically based on their correspondence to the ITC classification image.Through this process, spurious treetops in the valley and canopy gaps were excluded as they did not correspond to a class.An example classification showing the distribution of treetops by species and relative differences in density is given in Figure 5

Generating Treetop Information for Forest Compartments
A count of the number of trees associated with each species within delineated forest stands was undertaken (e.g., Figure 6) from which a stand map representing dominance of species and the density of stems was generated.
The mapping can be used to support forest management.In the example (Figure 7(a,b)), the mapping highlights stands that were dominated by C.obtusa, which has a higher timber value compared to others within stands.The mapping can be used to inform on actions such as thinning.A map of tree positions within each of three strata is given in Figure 7

Verification Based on Field Investigations
A comparison of the locations of the stems of C. obtusa trees identified using the tree top counting technique within seven forest compartments with those mapped through field survey indicated a close correspondence (for those ≥8 cm DBH) (Table 3).Three counts of trees were compared against the field data; treetops identified using local maxima filtering within the smoothed image, crowns identified using the ITC delineation and both treetops and crowns identified using a combined technique.The number of stems ranged from 399 to 1,065; treetops, from 647 to 1,381; crowns, from 201 to 387; and treetops generated by the new technique, from 416 to 938.
The treetop counts were more than that of the field stems, whereas the crown counts were less than that of the field stems.The error rate between the treetop counts and the stems counted in the field ranged from 22.4% to 62.2%.The error rate between the crown counts and the stems counted in the field ranged from −38.6% to −74.5%.On the contrary, the error rate between the numbers of treetops extracted with the new technique (NT) and stems from the field investigation ranged from 2.3% to −15.8%, showing that this method leads to improved accuracy over treetop and crown delineation used separately in stands 1, 3 and 7.In stands 2, 4, 5, and 6, the number of new technique treetops was slightly lower than that of stems from field survey data.These data indicate that it is difficult to extract treetops using image analysis, because these are two-storied stands with upper-story trees of L. kaempferi and P. densiflora and understory trees of C. obtusa and broad-leaved trees, which occurred as small-diameter crowns under the upper tree crowns.Stand 7 had the lowest error rate (2.3%), probably because the sharp, conical crowns of the dominant upper-story trees, C. obtusa and C. pisifera, were easier to extract using image analysis compared with crowns of other tree species [16].
Table 4 lists the numbers of stems and tree crowns, and the results of the New Treetop Technique by tree species (NTT).Generally, the number of crowns from the image analysis was fewer than that of stems from the field survey, primarily because some trees in the image were misidentified as one large tree crown rather than a cluster of trees.However, the new image analysis yielded treetops closer to the number of stems from field survey.Stem errors were derived from these data for accuracy verification.
Two types of stem errors were derived from these data for accuracy assessments, the later one (NTT-error 2) excluding understory stems.From the field survey, 118 of the 399 trees in stand 1 were L. kaempferi, and 252 of the 837 trees in stand 2 were P. densiflora.Of the 684 trees in stand 3, 186 were L. kaempferi and 152 were P. densiflora.In stands 4, 5, 6, and 7, C. obtusa and C. pisifera were dominant, with numbers of trees ranging from 352 to 554.Of the numerous broad-leaved trees, the majority was understory and intermediate trees; the numbers of broad-leaved understory trees ranged from 26 to 252 in all stands.The numbers of planted C. obtusa trees ranged from 29 to 194 in stands 2, 4, 5, 6, and 7, which were two-storied plantations.Planted C. japonica trees were mainly understory trees in stands 3 and 5 with 73 trees and 35 trees.
Compared with the respective numbers of stems from the survey, the numbers of C. obtuse, C. pisifera, and broad-leaved treetops determined using the new technique were lower, and the numbers of P. densiflora and L. kaempferi treetops were higher.
The error rates for treetops by tree species between NTT and stems from the field investigation are shown as NTT errors in Table 4.The error rates for C. obtusa and C. pisifera ranged from −16.0% to −89.6%, except in stands 1 and 3, where fewer of these trees grew.The error rates ranged from 15.7% to 183.6% for P. densiflora, from 11.9% to 487.1% for L. kaempferi, and from −33.3% to −100.0% for C. japonica.For the broad-leaved trees, the error rates ranged from −19.0% to −82.4%.
In stands 2, 4, 5, 6, and 7, there were fewer C. obtusa and C. pisifera treetops compared with the numbers of stems from the field survey because many of these were understory trees.However, there were more P. densiflora and L. kaempferi treetops compared with the numbers of stems because, owing to their large crowns, one crown was sometimes misidentified as a cluster of crowns.The error rate of the new technique, excluding understory stems, was obviously improved, as shown by NTT error 2 in Table 4.The error rates for total numbers of C. obtusa and C. pisifera trees per stand, excluding understory stems, ranged from −6.2% to −55.8%, and error rates for broad-leaved trees ranged from 8.2% to 63.1%.The results for multilayered or two-storied stands were inferior to those of single layer stands or pure conifer plantations because multilayered stands have some intermediate or understory trees that are difficult to extract using image analysis.

Figure 1 .
Figure 1.Treetops and tree crowns.(a) Trees seen from the side.(b) Trees seen from the air.

Figure 4 .
Figure 4.A texture homogeneity measure criteria is used to separate small tree areas from big tree.
(b).Patches of C. obtusa, P. densiflora, and L. kaempferi were clear, indicating a good representation of conifer plantations in the area.

Figure 5 .Figure 6 .
Figure 5. Identification of tree species (a) Treetops as red crosses overlaid on the individual tree-crown (ITC) species classification.(b) Treetops species identified using a combined ITC approach and treetop extraction method. (b).

Figure 7 .
Figure 7. (a) Species and stem locations, with C. obtusa trees indicated by large red points overlaid on true color composite image.(b) Map of tree positions within each of three strata(upper, intermediate and understory) from field data.

Table 2 .
Forest resource conditions of the seven compartments in the study area at Shinshu University Campus Forest.

Table 3 .
Comparison of ground survey data with image analysis of the seven compartments.Stems: Count of trees from field survey; Tree tops: Count of tree tops from image analysis; Crowns: Count of Crowns from image analysis; New tech.: Count of treetops using new technique from image analysis; TT-Error.: (Treetops − Stems) ÷ Stems × 100; CR-Error.: (Crowns − Stems) ÷ Stems × 100; NT-Error.: (New tech.− Stems) ÷ Stems × 100.

Table 4 .
Comparison between ground survey and the new technique data by species in the seven compartments.