int. j. remote sensing, 2000, vol. 21, no. 1, 61±79 Incorporating texture into classi®cation of forest species composition from airborne multispectral images S. E. FRANKLIN², R. J. HALL³§, L. M. MOSKAL², A. J. MAUDIE² and M. B. LAVIGNE¶ ²Department of Geography, The University of Calgary, Calgary, Alberta, Canada T2N 1N4 ³Canadian Forest Service, Natural Resources Canada, Northern Forestry Centre, Edmonton, Alberta, Canada T6H 3S5 ¶Canadian Forest Service, Natural Resources Canada, Atlantic Forestry Centre, Fredericton, New Brunswick, Canada E3B 5P7 (Received 17 February 1998; in ®nal form 2 February 1999 ) Although research with digital airborne remote sensing data has been undertaken in di€erent ecoregions to classify forested areas, the potential role of such imagery in deriving information to assist forest management has not yet been fully de®ned. The objective of this study was to determine the extent that the addition of texture could improve spectral classi®cation of high spatial resolution images (pixel size < 1m). These images represented pure and mixedwood forest stands from ecoregions in Alberta and New Brunswick, Canada. This study employed a judicious, selective application of texture to stands within a hierarchical classi®cation framework. In Alberta, the addition of texture made a modest improvement in classi®cation accuracy from 60% to 65%. In New Brunswick, the application of texture to selected land cover types resulted in an overall 12% improvement in classi®cation accuracy. The addition of image texture increased classi®cation accuracy for high spatial detail imagery relative to low spatial detail imagery. Incorporating texture into classi®cation also improved classi®cation accuracies for hardwood stands more so than for softwood stands, but greater attention to stand structure and composition will be needed in future work. Classi®cation accuracies on the order of 60±65% were achieved with simple texture derivatives, maximum likelihood decision rules and conventional classi®cation methods. Abstract. 1. Introduction The extent to which stand structure can be detected, classi®ed and mapped will determine the informational value of classi®ed airborne images in the inventory and management of the world’s forest resources. A forest stand is characterized by several attributes that include species composition, crown closure, height and age class. Examples from using remote sensing of airborne images are required to develop an understanding of the roles such images and methods can play in forest inventory, resource database development, forest management and modelling. These examples §Author to whom correspondence should be addressed. International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online Ñ2000 Taylor & Francis Ltd http://www.tandf.co.uk/journals/tf/01431161.html 62 S. E. Franklin et al. must be linked to existing forest inventory systems, which are stand descriptions consisting of the same structural attributes (i.e., species, crown closure, height) that are currently derived from the manual interpretation of aerial photographs (Leckie et al. 1995, Magnussen 1997). It is generally acknowledged that digital remote sensing can provide information that is not currently part of an existing forest inventory. One view that limits the use of remote sensing in forestry, however, is that results of digital image analysis should be comparable, or even superior, to the inventory results generated from existing analogue methods. In addition, the image analysis methods employed must be relatively simple, well-understood and readily accessible to managers responsible for development and implementation of resource management plans within a framework of ecological, socio-economic and environmental considerations. The premise of this study is that e€ orts toward an ecosystem-based approach to resource management, may be aided by the increasing availability of image processing software on the `desktop’ that can be easily integrated with geographic information systems (GIS). As Leckie et al. (1995: p. 337) have pointed out: `Use of digital high-resolution (< 1 m) multispectral imagery as an alternative to aerial photography for forest inventory mapping is a possible revolutionary innovation’. Previous studies in Canadian forests (Treitz et al. 1985, Smith et al. 1991, Franklin et al. 1991, Heygi et al. 1992, Franklin and McDermid 1993, Eldridge and Edwards 1993, St-Onge and Cavayas 1995, 1997) and elsewhere (e.g., Baulies and Pons 1995), have indicated that digital image processing methods applied to airborne images can aid in forest inventory, but that much e€ ort is required before these procedures can replace or even replicate existing inventories for a wide variety of reasons. In the long term, it is also now apparent that more widespread use of digital remote sensing may not occur simply because new applications, such as an assessment of stand productivity, can be accomplished. For example, Leaf Area Index (LAI) is not currently considered part of the existing forest inventory in Canada, and yet is an important structural attribute of ecosystems and forest stands that is related to productivity. Is the lack of LAI in the forest inventory a result of the fact that LAI cannot be estimated using air photo interpretation methods? Or, that LAI represents an aspect of the forest not directly related to timber volume, a basic inventory attribute in Canadian forestry? The increasing demands for new, timely and reliable information, such as LAI estimates, results in greater opportunities to develop new remote sensing applications. Forest inventory originates with a stand discrimination or classi®cation strategy that is necessarily regional and narrow in scope (Leckie and Gillis 1995). In Canada, this scope is de®ned provincially, and then combined nationally to provide annual perspectives on the status of forest resources in Canada (Natural Resources Canada 1997, 1998). Each province has established an inventory classi®cation system (Gillis and Leckie 1993) that results in mapped areas represented by polygons in a GIS database. It is this inventory standard, or this de®nition of stand strata, that digital remote sensing must attempt to emulate in a wide range of forest conditions. Only then will more e€ ort be justifed to resolve some of the more obvious problems in a full-scale adoption of the digital environment (Ryerson 1989, Pitt et al. 1997). The onus is clearly on the remote sensing community to provide convincing examples and illustrations of the digital use of airborne images in forest inventory applications. To date, the common approach to remote sensing for forest inventory is to apply an empirical classi®cation or estimation technique. The core of this Incorporating texture into forest classi®cation 63 method involves establishing an empirical relationship between spectral measurements and ®eld estimates of inventory variables, such as species or density, and the subsequent extension of these relationships to larger areas by a classi®cation decision rule. Recent studies have focused on extracting information from the image prior to developing classi®cation rules (Hay and Niemann 1994), or, using ancillary data such as a digital elevation model (DEM) to develop decision rules for forest classi®cation (Franklin 1994). Several studies (Fournier et al. 1995, St-Onge and Cavayas 1995, 1997, Lark 1996), have reported that texture analysis can improve classi®cation accuracy, but much work remains on how and where texture analysis can be e€ ective. Texture derivatives (Roach and Fung 1994, Wulder et al. 1998) and texture segmentation methods (Lobo 1997, Ryherd and Woodcock 1997), for example, require further development and testing in forestry and ecological mapping applications. The objective of this study was to evaluate the use of two texture measures when combined with spectral features in the classi®cation of forest structure as de®ned by provincial forest inventory procedures. Of interest was to classify a range of stands from pure coniferous and deciduous to mixed-wood stands in southern Alberta and south-eastern New Brunswick from high spatial resolution airborne images. A range of forest stands in di€ erent ecological conditions was selected to provide the widest possible range of test conditions for this initial study. Texture information represents the spatial variation in image tone (i.e., digital grey values) that is the result of the arrangement of forest vegetation and other objects in a digital image. The variability in stand structure results in unique variations of image tones that can be used to stratify stands, and to possibly increase the accuracy of forest classi®cation and mapping of biophysical attributes. 2. Study area and data collection The forests in this study are located adjacent to Kananaskis Country Provincial Park in Alberta, and in the Fundy Model Forest in south-eastern New Brunswick (®gure 1). The Alberta site is considered part of the Canadian Montane Forest Region (Strong and Leggat 1981). The relatively well-drained lower slopes of alluvial fans and lacustrine deposits are dominated by aspen (Populus tremuloides Michx.) and the upper, dryer slopes are dominated by lodgepole pine (Pinus contorta Dougl.). White spruce (Picea glauca (Moench) Voss) and balsam poplar (Populus balsamifera L.) are usually found in stands dominated by pine and aspen, respectively. The New Brunswick site is considered part of the Canadian Acadian Forest Region (Rowe 1972). Northern tolerant hardwoods, sugar maple (Acer saccharum Marsh.), yellow birch (Betula alleghaniensis Britton) and beech (Fagus grandifola Ehrh.) are common on upper slopes and ridges. Conifer stands composed of balsam ®r (Abies balsamea (L.) Mill.), red spruce (Picea rubens Sarg.), white spruce and black spruce (Picea mariana (Mill.) BSP) are most common on lower slopes, and mixedwood stands are frequently found on mid-slopes. Red maple (Acer rubrum L.), trembling aspen, bigtooth aspen (Populus grandidentata Michx.) and white birch (Betula papyri®ra Marsh.) are common in large part because of frequent harvesting. Jack pine (Pinus banksiana Lamb.) frequently occurs on areas with sandy soils, and both red pine (Pinus resinosa Ait.) and white pine (Pinus strobus L.) are found in hardwood and mixed-wood stands. White ash (Fraxinus americana L.), eastern white cedar (T huja occidentalis L.), red oak (Quercus rubra L.) and eastern hemlock (T suga canadensis (L.) Carr.) are found in the region. Stands with numerous species in the 64 S. E. Franklin et al. Figure 1. Location of study areas in boreal forests in south-western Alberta and southeastern New Brunswick. overstory (four or more) are common in this region because of the large number of tree species found in abundance. 2.1. Field data collection Table 1 contains a listing of the ®eld data characterizing each of the 26 Alberta plots, and table 2 contains a similar listing of the ®eld data characterizing each of the 17 New Brunswick plots used in the analysis. Field plots that were 10 mÖ 10 m in size were established in Alberta and 20 mÖ 20 m in New Brunswick to represent a range of conditions on near-level terrain to minimize topographic e€ ects. Standard Alberta Vegetation Inventory (AVI) methods (Alberta Forestry, Lands and Wildlife 1991), and New Brunswick Integrated Land Classi®cation System (New Brunswick Department of Natural Resources 1996) methods were applied to classify each plot established in the Alberta and New Brunswick study areas, respectively. Estimates of crown closure in Alberta were made at ®ve locations within each plot with the aid of a hand-held spherical densiometer. Plot tree heights were taken for two dominant or co-dominant trees with a hand-held clinometer. Diameter at breast height (DBH) was also measured for all trees in a plot, and extensive notes about understory species were collected. Field measurements of individual trees in each plot were used to generate a stand descriptor label that included (Gillis and Leckie 1993): 1. 2. 3. 4. Species (i.e., percent canopy for dominant and co-dominant trees) Density (i.e., four or ®ve categories related to crown closure) Height Stand origin (corresponding to year of origin in Alberta) or age (corresponding to young, immature, mature or overmature in New Brunswick) and 5. Other attributes such as understory, site quality and drainage. Forest inventory classi®cations were generated in the ®eld and then compared to Table 1. Summary of ®eld data for the 26 Alberta forest plots. Species composition according to the AVI code Plot ID Middlestory Merged Pl5 Pb5 Aw10 Aw10 Pl10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw5 Pb4 Sw1 Aw9 Pl1 Pl10 Pl3 Aw3 Sw2 Pb2 Aw7 Sw3 Pb4 Aw3 Pl2 Sw1 Pl10 Sw5 Pl3 Aw3 Pl6 Sw4 Pl10 Pl4 Sw3 Aw2 Pl1 Pl5 Sw5 Pl6 Sw4 Pl8 Sw2 Pl8 Sw2 Pl10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw10 Aw4 Pb4 Sw2 Aw8 Pl2 Aw7 Pl3 Aw3 Pb3 Pl3 Sw1 Sw4 Aw4 Pb2 Pb4 Sw3 Aw2 Pl1 Aw5 Pl4 Sw1 Sw4 Pb4 Aw1 Pl1 Sw4 Pl3 Pb3 Pl7 Aw3 Sw4 Pl4 Pb1 Aw1 Sw5 Pl4 Aw1 Pl5 Sw5 Pl7 Sw3 Pl7 Sw3 Pl10 70 70 70 64 70 70 64 70 70 64 70 64 64 70 68 64 64 68 64 64 64 64 66 73 73 64 Aw10 Aw10 Pb5 Sw5 Aw5 Pl4 Pb1 Aw9 Pl1 Pb7 Sw3 Sw5 Pb5 Sw10 Aw6 Pl3 Sw1 Pb7 Sw3 Sw5 Pb5 Aw6 Pl4 Pl5 Sw3 Aw1 Pb1 Sw5 Pl3 Aw2 Sw10 Pl5 Sw5 Sw10 Pl6 Aw2 Pb2 Crown closure³ Percent AVI code 46 49 36 54 64 53 36 52 55 51 54 60 50 65 59 60 61 63 54 52 63 60 30 50 50 38 B B B C C C B C C C C C B C C C C C C C C C A B B B ²Species composition: Aw = Trembling Aspen, Pb = Balsam Poplar, Pl = Lodgepole Pine, Sw = White Spruce, 1= 10%, 2= 20%, ... 10= 100%. ³Crown closure: A= 7±30%, B= 31±50%, C= 51±70%, D= 71±100%. §AVI Inventory polygon descriptions. Polygon no. 64 66 Polygon no. B11Pl10 A10 Pl7Sw2Aw1 B11Pll4Aw4Sw2 A6Pl5Aw4Sw1 B15Pl8Aw2 B10Pl7Aw2Sw1 70 73 AVI code* overstory middlestory overstory middlestory *AVI code is de®ned as crown closure³, stand height to closest metre, and species composition². C12Aw9Pl1 A4Sw8Aw2 C17Pl5Aw5 A10Sw8Pl2 65 68 overstory middlestory overstory middlestory overstory middlestory AVI code* Incorporating texture into forest classi®cation 14 21 22 10 19 20 23 24 25 26 17 18 9 13 15 16 8 12 6 7 11 4 1 2 3 5 Overstory² AVI inventory polygon number§ 66 S. E. Franklin et al. Table 2. Summary of ®eld data for the 17 New Brunswick plots. Species composition² Density (stems haÕ 1 ) LAI New Brunswick inventory polygon label³ HB1 HB3 HB4 HB5 HB6 HB7 HB8 HB9 HB10 jP1 0 rM6 bF2 jP2 rM4 rS3 wP2 jP1 wB1 wB8 rM2 rM5 wB4 Be1 rM5 wB2 rS2 bF1 rM5 wB5 rM3 bF3 wB2 wP2 jP6 rM2 gB1 tA1 5000 295 1350 1300 1275 525 625 625 1825 2.63 3.86 6.77 4.76 4.71 3.89 3.84 3.84 6.01 jP1 0 IH4 SP3 SW2 SP7 WP2 IH9 IH9 IH4 SP3 WP3 TH4 IH3 SW3 SP4 WP4 IH7 WP2 DU2 DU3 DU5 DU6 DU7 DU8 DU9 DU10 jP1 0 wS1 0 rM5 yB2 Be1 sM1 wB1 bF4 rM4 rS1 wB1 Be3 wB3 rM2 rS2 wB6 rM2 rS2 rM7 rS1 tA1 wB1 wB7 rM3 2500 1475 1675 1050 1000 1375 675 1275 5.61 5.83 4.93 5.87 6.88 5.25 3.28 4.30 jP1 0 SP1 0 IH6 TH4 IH4 TH4 SW2 TH1 0 IH6 TH3 IH6 TH3 IH6 TH3 Plot ID ²See ®gure 3 for a description of species composition labels. ³Species descriptors: IH= tA, wB, gB; TH= yB, Be, rM, sM; SW= mixed pine, spruce, ®r; SP= wS, rS. the GIS forest inventory database. In every case, the ®eld inventory was more detailed than the existing forest inventory. This is attributed to the photo interpretation process that results in an average descriptor for the stand. For example, for the 26 plots surveyed in the ®eld in the Alberta site, only ®ve AVI polygons were mapped in the GIS (table 1). The more detailed ®eld labelling was retained for this analysis because of the requirement to compare the image classi®cation with its corresponding stand type in the ®eld. The accuracy by which the image data could predict pixel membership within the plots was then determined by comparison with the ®eld measurements. 2.2. Airborne remote sensing data collection The high spatial resolution data included airborne multispectral video images acquired July 1996 in Alberta (Roberts 1995, Gerylo et al. 1997), and airborne spectrographic images acquired August 1995 in New Brunswick (Anger et al. 1994, Wulder et al. 1996). The spectral data in Alberta consisted of three image bands from 490 to 850 nm. In New Brunswick, the spectral data included ®ve image bands from 560 to 750 nm. Image acquisition for both study areas were accomplished by a light aircraft ¯ying perpendicular to the major solar axis, and was restricted to high solar angle times on days with little or no cloud cover. Atmospheric correction procedures included a dark-pixel object subtraction routine, using the deep waters of nearby lakes, and an empirical calibration using pseudo-invariant objects measured with the ASD Personal Spectrometer II during each airborne mission (Milton et al. 1997). The highest spatial resolution images used in this study approached 0.3 m. In Alberta, multispectral video (MSV) (Roberts 1995) images were obtained from three Incorporating texture into forest classi®cation 67 co-registered, ®ltered and calibrated SONY XC-75 CCD (charge coupled device) video cameras. Six short ¯ight lines were ¯own 200 m above the site, and at ®ve increasing altitudes up to 3500 m, resulting in images with approximately 0.3, 1, 2, 4, 6 and 8 m spatial resolutions, respectively. In New Brunswick, images were obtained from the compact airborne spectrographic imager (CASI) (Anger et al. 1994) over two 10-km long ¯ight lines, from an altitude of approximately 3000 m that resulted in a spatial resolution of 1 m. The sample plots were located on the geocoded images by a di€ erential GPS ground survey and a GPS base station. Visual clues were also used to locate plots when problems with overstory de¯ection of the radio signal (Ho€ man-Wellenhof et al. 1992) were suspected. Plot locational accuracy was within 1 m of its true location, and plots were generally restricted to near-nadir positions and digital image frame centres to reduce bidirectional e€ ects, lens vignetting and radial displacement. None of these potential sources of error appeared to compromise the image samples used in the study when subjected to detailed image interpretation on computer displays. 3. Methods 3.1. Classi®cation hierarchy The relationships between the image data and plot-level forest inventory were obtained by image classi®cations using the maximum likelihood decision rule. The discriminating power of the spectral and textural variables was determined by the success of the classi®cation shown in contingency tables and per cent classi®cation accuracy. An analysis of accuracy was based on a series of classi®cation tests whereby the results of pixel class membership were compiled for di€ erent levels of inventory class generalization. Classes at each level of the hierarchy were either accepted, eliminated, merged or modi®ed according to the structure of the forest inventory system, and based on the maximum detail that could be retained in the pixel classi®cations for that plot. This process provided insight into the use of texture as a selective variable that is class speci®c (Hay and Niemann 1994, Ryherd and Woodcock 1997). The ®rst level of the classi®cation hierarchy assumed that each plot in the sample was a separate and distinct class. The second level assumed that similar plots represented the range of conditions in a single class described by a species composition label. The third and fourth levels of the classi®cation hierarchy represented increasing generalizations of the forest inventory description. In the New Brunswick sample, the spectral classi®cation accuracy in each original plot was examined to determine whether to include texture into classi®cation for that plot, or to accept the accuracy attained without further discrimination. In conventional classi®cation there is the assumption that every pixel in the plot is a member of the class that it represents (Franklin 1994, Hay and Niemann 1994, Fournier et al. 1995). This can be a source of pixel classi®cation error because many pixels in any given plot will not uniquely belong to the class the plot is supposed to represent. Earlier work (Franklin 1994) has shown that a composite spectral signature consisting of many objects contributing to re¯ectance can be used to classify high resolution image data with a reasonable (i.e., nonrandom) degree of accuracy. The use of image texture in the hierarchical classi®cation of forest inventory plots should provide additional discriminatory power because, while the spectral response is an aggregate of the individual objects contributing to re¯ectance, the texture variable 68 S. E. Franklin et al. represents the spatial pattern in which these objects occur. The hypothesis that adding texture to spectral classi®cation will increase land cover map accuracy is consistent with the use of texture as an element of photo interpretation in forest inventory (Leckie et al. 1995). The classi®cation accuracy in this study is reported as the percentage of correct pixels in the ®eld plot that were classi®ed into the class from which they originated, as determined by the ®eld survey. 3.2. Image texture measures Image texture is a quanti®cation of the spatial variation of image tone values that are related to changes in the spatial distribution of forest vegetation, both in the vertical and horizontal dimensions. Image texture is a complex concept and application. Simple texture measures may be derived by comparing the values of the digital numbers within a window, or in derivatives of a matrix representing the greylevel co-occurrence within the window (Haralick 1986, Oja and Valkealahti 1996). Two co-occurrence texture derivatives described by Franklin and Peddle (1987) were applied to the infrared channel (approximately 740±760 nm) to characterize the texture of the vegetation for each plot. Only simple texture measures were derived based on preliminary interpretations of image displays (Franklin and Peddle 1990) to illustrate their use and the degree of improvement possible in plot classi®cation. More complex multidimensional texture measures are possible, and subsequent study may attempt to optimize the texture measures for this application. The size of the window over which texture measures are derived may be altered or adapted to better represent the characteristics of the local area (Haralick 1986, Marceau et al. 1990, Ryherd and Woodcock 1997). Large windows (eg., 9 Ö 9 compared to 3 Ö 3) can provide better estimates of distributions and decrease the random error, but may encompass more than one stand type that could introduce systematic error. An increased ability to estimate forest stand parameters such as LAI, stand density and volume have been reported after customized windows sizes were used rather than ®xed, arbitrary windows (Franklin and McDermid 1993, Wulder et al. 1998). These are data-driven geographic windows, rather than arbitrary, ®xed geometric windows (Franklin et al. 1996). Image semivariance, based on image variograms, was used to relate pixel self-similarity over a distance to suggest optimal geographic window sizes based on the characteristics of a particular spectral bandset. In images of lower spatial resolution with fewer pixels than the optimal window size predicted by the semivariogram, the window was restricted to the plot area. An inter-pixel sampling distance of one and a zero degree angle were utilized on 8-bit linearly transformed 16-bit data to calculate the texture derivatives from the spatial co-occurrence matrix. Texture measures homogeneity and entropy were used in this study (PCI Inc. 1994): n m Homogeneity = ž ž j -1 i-1 n P(i,j ) (1 + [R (i )Õ C( j )]2 ) (1) m Entropy = ž žP (i,j ) 1n (p(i,j )) if P(i,j )^0 otherwise, Ebtropy = 0 (2) j -1 i-1 where P(i,j ) is the spatial co-occurrence matrix element, R (i ) is the grey level value for a row, and C( j ) is the grey level value for a column. 69 Incorporating texture into forest classi®cation 4. Results 4.1. AVI classes Based on spectral and texture classi®cations of the images acquired from six di€ erent altitudes in Alberta, the highest spatial resolution images were the least accurate at each level of the hierarchy. These images improved the most, however, when texture derivatives were added to the classi®cation decision rule (®gure 2, table 3). In the ®rst level of the hierarchy over all the classes, spectral data alone yielded just 15% accuracy at the highest level of spatial detail; this improved to 40% with the higher altitude (lower spatial resolution) imagery. At the second and third levels of the hierarchy, the same pattern was found: 23% to 34% accuracy, and then, 52% to 68% accuracy, respectively. These patterns con®rmed that the mean spectral response measured in the plot did not produce an accurate classi®cation of the pixels in that plot alone, but the addition of simple texture derivatives improved the overall relationship between image data and ®eld observations. When the six spatial resolutions for a given level of the classi®cation hierarchy were averaged, the airborne images were 26% correct at the highest level of detail in the classi®cation hierarchy (®gure 2). This accuracy was reduced to 22% when only texture variables were used, and improved to 37% when both texture and spectral variables were used in the decision rules. The second level of the hierarchy showed a di€ erent pattern of 28% to 34% to 46% (®gure 2). The third level of the hierarchy showed a similar pattern of 60% to 57% to 65%. At this level of classi®cation detail, the texture variables were more powerful than the spectral variables in discriminating the inventory classes at four of the six spatial resolutions examined in the study. The exceptions included the two images with the lowest spatial resolutions. On average, the spectral discrimination accuracy approximated or exceeded the texture discrimination accuracy. The range of accuracy was large for di€ erent levels of the hierarchy and for di€ erent image resolutions. The highest accuracy found in the classi®cations was produced by the texture data at the highest spatial resolution, but with the lowest level of detail in the class hierarchy (97%, three classes). The lowest accuracy found in the classi®cations was produced by the texture data at the lowest spatial resolution but with the highest level of detail in the class hierarchy (10%, 26 classes). Table 3. Summary of the classi®cation results for two di€erent spatial resolution images and the mean of all images at each level of the hierarchy over the Alberta plots. Class structure² Level 1 Level 2 Level 3 Spatial resolution (m) Spectral data alone Texture data alone Combined 0.3 8.0 mean³ 0.3 8.0 mean 0.3 8.0 mean 15 40 26 23 34 28 52 68 60 23 10 22 32 17 34 76 41 57 31 44 37 44 40 46 74 64 65 ²See ®gure 2 for a description of species composition labels. ³Mean of all six spatial resolutions at each level of the hierarchy shown in box at top of ®gure 2 for each set of plot results. 70 S. E. Franklin et al. Figure 2. Spectral response, texture derivative, and combined spectral and texture hierarchical classi®cation for 26 Alberta (AVI) forest plots. Incorporating texture into forest classi®cation 71 Individual class performance in the classi®cation varied greatly by spatial resolution and level of classi®cation hierarchy. A few notable patterns emerged, however, which appeared to strengthen the rationale underlying the use of image texture in a selective method, rather than application to every class and training sample. For example, the aspen stand in plot 26 varied in accuracy from 68% at the lowest spatial resolution in the ®rst level of the hierarchy, and was reduced in accuracy to 45% with the addition of texture variables. The original 68% accuracy attained is considered acceptable, showing that this plot is spectrally distinct for image classi®cation purposes. Texture derivatives appeared to confuse this distinctiveness and reduced the accuracy achieved with the spectral data. This ®nding was not consistent with the satellite image texture segmentation results reported by Ryherd and Woodcock (1997), where they found the combination of texture and spectral data never degraded segmentation accuracies. Classi®ers may be more sensitive, however, to texture variables than are segmentation routines. Plot 25, for example, contained a similar stand of aspen with an increased density (table 1) that was improved in accuracy from 26% to 77% with these same image data. This plot was more similar to the overall pattern that was dominated by increased discrimination provided by texture derivatives, particularly in the mixed-wood stands. 4.2. New Brunswick Department of Natural Resources (NBDNR) classes A classi®cation hierarchy with a single image dataset was also constructed for the New Brunswick study plots (®gure 3). Texture derivatives provided only marginal increases in classi®cation accuracy relative to using spectral data alone. At the ®rst level of the hierarchy, accuracy was 42% using spectral data alone, 23% using texture, and 42% using the combination of spectral and texture data in the decision rule (®gure 3). The addition of texture increased accuracy consistently across the inventory classes for the middle two levels of the hierarchy, but the levels of accuracy were still relatively low. At level four of the hierarchy, accuracy improved to 57% with spectral and texture data. Only a small 2% increase in accuracy was achieved in comparison to using spectral data alone. The addition of texture measures to the classi®cation resulted in less discrimination among stands of similar species composition in comparison to use of only the spectral data. The relative frequency and dominance of 12 hardwood stands relative to ®ve softwood stands, and its variable composition may explain, in part, the results attained (®gure 3). Hardwood stands were adequately classi®ed with the spectral data and little additional discrimination was achieved with the addition of texture. In this study, however, texture refers to the co-occurrence texture used in the study, of which many formulations of texture were possible and could be useful to describing the textural characteristics of a given land cover class. A second classi®cation sequence was applied to the New Brunswick plots, whereby texture was not applied to spectral classi®cations that achieved 60% accuracy, or to classes where texture made little or no improvement to class discrimination (®gure 4). The overall classi®cation accuracy did not change signi®cantly with 54% (®gure 4) compared to 57% (®gure 3), but more importantly, the number of classes retained in the class structure equalled the original level of the hierarchy. When compared to the original spectral classi®cation at level 1 of the hierarchy (42%), a 12% increase in accuracy was achieved by using texture selectively in comparison to applying texture to all land cover classes. The presence of texture variability was illustrated for a pure jack pine plantation 72 S. E. Franklin et al. Figure 3. Spectral and texture hierarchical classi®cation for New Brunswick (NBDNR) forest plots. (®gure 5 (a), plot DU2) in comparison to a plot comprising jack pine, red maple, grey birch and trembling aspen (®gure 5 (b), plot HB10). Plot DU2 (®gure 5 (a)) did not increase in accuracy with the addition of texture and was adequately classi®ed using spectral data alone (87%). This plot was a single species, closed canopy Incorporating texture into forest classi®cation 73 Figure 4. Selective texture classi®cation for New Brunswick (NBDNRI) forest plots. plantation jack pine that presented a very homogeneous spectral response to the sensor. Consequently, the mean spectral response and standard deviation were suciently unique (compared to other classes in the sample) that a relatively high classi®cation accuracy could be achieved. Plot HB10 (®gure 5 (b)) increased in accuracy with the addition of texture from a low of 15% using spectral data, to a 74 S. E. Franklin et al. (a) (b) Incorporating texture into forest classi®cation 75 maximum of 45% using spectral plus texture data in the decision rule (®gure 4). This plot contained a mixture of jack pine and at least three di€ erent deciduous species in a closed canopy. Visually this plot was inherently `textural’ and although the improvement in classi®cation accuracy was impressive, the relatively low accuracy even after adding texture suggests alternative texture measures may be required to discriminate this inventory class (Roach and Fung 1994). 4.3. Discussion The highest overall classi®cation accuracies reported in these analyses of multispectral, high spatial resolution, airborne videographic and spectrographic images in Alberta and New Brunswick were 65% and 57%, respectively. These results were derived from classi®cation of spectral and texture features, and were compared to ®eld plots inventoried according to their prevailing provincial systems. The spectral classes were highly detailed and similar to the species composition labels used in forest inventory operations. The accuracies achieved were likely comparable to those that were obtained in the present GIS-based, forest inventory derived primarily through aerial photo mapping. The classi®cation accuracy results obtained with this study were similar to other studies that have incorporated texture analysis in land cover mapping (Franklin and Peddle 1990, Ryherd and Woodcock 1997). Texture processing generally improves classi®cation accuracy in the order of 10±15%. There were few comparable applications of airborne image texture analysis in forest inventory studies. Earlier work by Franklin and McDermid (1993) reported the best overall classi®cation accuracy based on airborne spectrographic images and texture derivatives was 63% in seven stand volume classes of Alberta lodgepole pine stands. Similar classi®cation methods and accuracy assessment procedures were used, and the original classi®cation accuracy using only three-band spectral data was 46%. Additional customized multiband texture and ®lter processing of the image data yielded 75% classi®cation accuracy. St-Onge and Cavayas (1997) reported error levels of approximately 20% for forest cover and density classes, but their approach incorporated texture and image segmentation in a more advanced image processing methodology (Lobo 1997). Similarly, Gougeon (1995, 1997) has described the results of classi®cation procedures, based on individual tree recognition algorithms, that has generated accuracies in the range of 68% in primarily pure conifer stands. In each of these studies, comparisons were made between digital image processing results and ®eld data or air photo interpretations that were considered 100% correct. A more reasonable perspective is to assume that error exists in both ®eld and remote sensing surveys, and that both datasets can lead to a reasonable generalization of the forest inventory. Direct comparisons of air photo interpretation and ®eld inventory assessments have also yielded similar error rates (Biging and Congalton 1991, Figure 5. Example infrared image data in two New Brunswick forest plots. The green colour shades represent conifer species, and the pink to red shades represent deciduous species. (a) Plot DU2Ða homogeneous jack pine plantation, closed canopy, high classi®cation accuracy (87%) based on spectral data alone. (b) Plot HB10Ða complex mixture of jack pine and three deciduous species that appears inherently textural, closed canopy, low classi®cation accuracy based on spectral data alone (15%) but increased accuracy when texture derivatives were added (45%). 76 S. E. Franklin et al. Fent et al. 1995, Hall and Fent 1996) to those reported from digital image processing studies. All of the above comparable studies incorporate texture into the analysis and all have achieved results that are unambiguous in con®rming the validity of the concept of digital image texture. The fact that digital image data can be classi®ed with an accuracy and con®dence level consistent with existing methods of forest inventory, however, does not imply stands can be mapped well (Franklin 1994). Clearly, di€ erent and ¯exible strategies are required for the incorporation of texture variables, as presented in this summary of the Alberta and New Brunswick airborne image classi®cation results, and in the reporting of these results to the literature. A checklist of the obvious sources of error or uncertainty in the application of digital image texture to forest inventory classi®cation can be identi®ed and include: 1. The choice of the texture measure itself (Roach and Fung 1994, Lark 1996, Wulder et al. 1998) 2. The area or window size for the texture calculation (Hay and Niemann 1994, Marceau et al. 1990) 3. The way in which the texture measure is used in the analysis (i.e., classi®cation, preprocessing, segmentation, combination of these, other) (Fournier et al. 1995, St-Onge and Cavayas 1995, Lobo 1997) 4. The type and range of inventory classes that are considered (Leckie et al. 1995, Ryherd and Woodcock 1997, St Onge and Cavayas 1997) 5. The way in which the comparison to ®eld data is made (i.e., classi®cation accuracy, regression, other) (Franklin and McDermid 1993, Baulies and Pons 1995, Fent et al. 1995) 6. The type of image data available, or the way in which the image acquisition characteristics (i.e., spatial resolution, spectral bands, ancillary information, and so on) are matched to the forest conditions under consideration (Milton et al. 1997, Wulder et al. 1996) 7. The positioning accuracy of a plot and the plot size especially in relation to image spatial resolution, and 8. The ultimate purpose of the classi®cation and the subsequent level of accuracy required (i.e., the interpretive reason to produce maps based on digital airborne remote sensing data and methods, or the accuracy levels required by the GIS database standards employed, or the context of the classi®cation such as update, inventory and exploratory mapping). 5. Conclusion Most studies in forest classi®cation that have utilized spectral and texture features have tended to apply these features to all land cover classes. The selective application of texture within a hierarchical classi®cation framework, however, resulted in higher overall accuracies than application to all land cover classes. The highest classi®cation accuracies achieved were in the range of 60±65% for pure and mixed-wood inventory classes in two widely variable forests from Alberta and New Brunswick. These stands were organized by species composition in di€ erent levels of generalization based on detailed ®eld surveys which were more complete than existing GIS-based inventory systems. For the two study areas, the addition of texture generally improved classi®cation accuracy for hardwood stands more so than for softwood stands. This result was di€ erent than what has been previously reported because the use of texture Incorporating texture into forest classi®cation 77 generally improves classi®cation of softwood stands due to their characteristic structure and shadows. Future studies will further examine the in¯uence of stand structure and other inventory characteristics on image texture derivatives and their impact on classi®cation accuracy. Although the results in this paper were consistent with those reported in other Canadian studies and elsewhere using similar data and methods, a more judicious use of texture appears justi®ed. This study is part of a larger programme directed at deriving a greater understanding of the role high spatial resolution, multispectral airborne images and classi®cation methods may have in forest inventory classi®cation. Acknowledgments This research was supported by the Natural Science and Engineering Research Council (NSERC) and the Canadian Forest Service. The Alberta image data were provided by Dr Arthur Roberts (Department of Geography, Simon Fraser University). The Alberta image database used in this study was created by Graham Gerylo and is gratefully acknowledged. References Alberta Forestry Lands and Wildlife, 1991, Alberta vegetation inventory standards manual. Alberta Environmental Protection, Resource Data Division, Edmonton, Alta. Version 2.1. Anger, C., Mah, S., and Babey, S., 1994, Technological enhancements to the compact airborne spectrographic imager (casi ). Paper Presented at the First International Airborne Remote Sensing Conference and Exhibition , Strasbourg, France, 11±15 September, 1994, Volume II (Environmental Research Institute of Michigan, Ann Arbor, Michigan), pp. 205±213. Baulies, X., and Pons, X., 1995, Approach to forest inventory and mapping by means of multi-spectral airborne data. International Journal of Remote Sensing, 16, 61±80. Biging, G., and Congalton, R., 1991, A comparison of photointerpretation and ground measurement of forest structure. Proceedings, 56th ASPRS Annual Convention, Vol. 3, Remote Sensing, Baltimore, Maryland (ASPRS: Bethesda, Maryland), pp. 6±14. Eldridge, N. R., and Edwards, G., 1993, Continuous tree class density surfaces derived from high resolution digital image analysis. Proceedings, GIS/L IS ¾93, Vancouver, BC (GIS World: Vancouver), pp. 947±952. Fent, L., Hall, R. J., and Nesby, R. K., 1995, Aerial ®lms for forest inventory: optimizing ®lm parameters. Photogrammetric Engineering and Remote Sensing, 61, 281±289. Fournier, R., Edwards, G., and Eldridge, N., 1995, A catalogue of potential spatial discriminators for high spatial resolution digital images of individual crowns. Canadian Journal of Remote Sensing, 21, 285±298. Franklin, S. E., 1994, Satellite remote sensing of forest type and landcover classes in the subalpine forest region, Kananaskis Valley, Alberta. Geocarto International , 7, 25±35. Franklin, S. E., and McDermid, G. J., 1993, Empirical relations between digital SPOT HRV and CASI spectral response and lodgepole pine (Pinus contorta) forest stand parameters. International Journal of Remote Sensing, 14, 2331±2348. Franklin, S. E., and Peddle, D. R., 1987, Texture analysis of digital image data using spatial co-occurrence. Computers & Geosciences, 13, 293±311. Franklin, S. E., and Peddle, D. R., 1990, Classi®cation of SPOT HRV imagery and texture features. International Journal of Remote Sensing, 11, 551±556. Franklin, S. E., Blodgett, C. F., Mah, S., and Wrightson, C., 1991, Sensitivity of casi data to anisotropic re¯ectance, terrain aspect and deciduous forest species. Canadian Journal of Remote Sensing, 17, 314±321. Franklin, S. E., Wulder, M. A., and Lavigne, M. B., 1996, Automated derivation of geographic windows for use in remote sensing digital image texture analysis. Computers and Geosciences, 22, 665±673. Gerylo, G., Franklin, S. E., Roberts, A., Milton, E. J., and Hall, R. J., 1997, Hierarchical Alberta Vegetation Inventory (AVI) classi®cation using aerial digital frame camera 78 S. E. Franklin et al. data. Proceedings, 19th Canadian Symposium on Remote Sensing (Canadian Aeronautics and Space Institute: Ottawa, Ontario), Paper number 4, Session E-7, pp. 7 CD-ROM. Gillis, M. D., and Leckie, D. G., 1993, Forestry Canada , Petawawa National Forestry Institute, Chalk River, Ontario. Information Report PI-X-114. Gougeon, F., 1995, A crown-following approach to the automatic delineation of individual tree crowns in high spatial resolution aerial imagery. Canadian Journal of Remote Sensing, 21, 274±283. Gougeon, F., 1997, Recognizing the forest from the trees: individual tree crown delineation, classi®cation and regrouping for inventory purposes. Presented paper, 3rd International Airborne Remote Sensing Conference and Exhibition , 7±10 July 1997, Copenhagen, Denmark (Environmental Research Institute of Michigan, Ann Arbor, Michigan), pp. 807±814. Hall, R. J., and Fent, L., 1996, In¯uence of aerial ®lm spectral sensitivity and texture on interpreting images of forest species composition. Canadian Journal of Remote Sensing, 22, 350±359. Haralick, R., 1986, Handbook of Pattern Recognition and Image Processing (London, Academic Press), pp. 247±279. Hay, G., and Niemann, O., 1994, Visualizing 3D texture: a three-dimensional structural approach to model forest texture. Canadian Journal of Remote Sensing, 20, 90±101. Heygi, F., Pilon, P., and Walker, P., 1992, Replacing aerial photos in resource inventories with airborne digital data and GIS. Proceedings, ASPRS/ACSM/RT I ¾92 T echnical Papers , Washington, DC, volume 5 (ASPRS: Bethesda, Maryland), pp. 510±516. Hoffman-Wellenhof, B., Lichtenegger, H., and Collins, J., 1992, Global Positioning Systems: theory and practice (Springer-Verlag: New York). Lark, R. M., 1996, Geostatistical description of texture on an aerial photograph for discriminating classes of land cover. International Journal of Remote Sensing, 17, 2115±2133. Leckie, D., and M. Gillis , 1995, Forest inventory in Canada with an emphasis on map production. Forestry Chronicle, 71, 74±88. Leckie, D., Beaubien, J., Gibson, J. R., O’Neill, N. T., Piekutwoski, T., and Joyce, S. P., 1995, Data processing and analysis for MIFUCAM: a trial of MEIS imagery for forest inventory mapping. Canadian Journal of Remote Sensing, 21, 337±356. Lobo, A., 1997, Image segmentation and discriminant analysis for the identi®cation of land cover units in ecology. IEEE T ransactions on Geoscience and Remote Sensing, 35, 1136±1145. Magnusson, S., 1997, A method for enhancing tree species proportions from aerial photos. Forestry Chronicle , 73, 479±487. Marceau, D., Howarth, P., Dubois, P., and Gratton, D., 1990, Evaluation of grey level co-occurrence matrix method for land-cover classi®cation using SPOT imagery. IEEE T ransactions on Geoscience and Remote Sensing, 28, 513±519. Milton, E. J., Lawless, K. P., Roberts, A., and Franklin, S. E., 1997, The e€ect of unresolved scene elements on the spectral response of calibration targets: an example. Canadian Journal of Remote Sensing, 23, 252±256. Natural Resources Canada, 1997, T he State of Canada’s Forests: learning from history. NRCan (Canadian Forest Service: Ottawa, Ontario). Natural Resources Canada, 1998, T he State of Canada’s Forests: the people’s forests. NRCan (Canadian Forest Service: Ottawa, Ontario). New Brunswick Department of Natural Resources, 1996, T he New Brunswick Integrated L and Classi®cation System (New Brunswick Department of Natural Resources: Fredericton, New Brunswick). Oja, E., and Valkealahti, K., 1996, Co-occurrence map: quantizing multidimensional texture histograms. Pattern Recognition L etters, 17, 723±730. PCI Inc., 1994, PACE radar analysis. V5.3, Richmond Hill, Ontario. Pitt, D. G., Wagner, R. G., Hall, R. J., King, D. J., Leckie, D. G., and Runesson, U., 1997, Use of remote sensing for forest vegetation managment: a problem analysis. Forestry Chronicle , 73, 459±477. Roach, D., and Fung, K., 1994, Fractal-based textural descriptors for remotely sensed forestry data. Canadian Journal of Remote Sensing, 20, 59±70. Roberts, A., 1995, Integrated MSV airborne remote sensing. Canadian Journal of Remote Sensing, 21, 214±224. Incorporating texture into forest classi®cation 79 Rowe, J. S., 1972, Forest regions of Canada. Environment Canada, Canadian Forest Service Publication No. 1300 (Ottawa, Ontario: Environment Canada). Ryerson, R. A., 1989, Image interpretation concerns for the 1990s and lessons from the past. Photogrammetric Engineering and Remote Sensing, 55, 1427±1430. Ryherd, S., and Woodcock, C., 1997, Combining spectral and texture data in the segmentation of remotely sensed images. Photogrammetric Engineering and Remote Sensing, 62, 181±194. Smith, N., Borstad, G., Hill, D., and Kerr, R., 1991, Using high-resolution airborne spectral data to estimate forest leaf area and stand structure. Canadian Journal of Forest Research, 21, 1127±1132. St-Onge, B., and Cavayas, F., 1995, Estimating forest stand structure from high resolution imagery using the directional variogram. International Journal of Remote Sensing, 16, 1999±2021. St-Onge, B., and Cavayas, F., 1997, Automated structure mapping from high resolution imagery based on directional semivariogram estimates. Remote Sensing of Environment , 61, 82±95. Strong, W., and Leggat, K., 1981, Ecoregions of Alberta (Alberta Department of Energy and Natural Resources: Edmonton). Treitz, P., Howarth, P., and Leckie, D. G., 1985, The capabilities of two airborne multispectral sensors for classifying coniferous forest species. Proceedings, 19th International Symposium on Remote Sensing of Environment (Environmental Research Institute of Michigan: Ann Arbor, Michigan), pp. 335±350. Wulder, M. A., Franklin, S. E., and Lavigne, M. B., 1996, High spatial resolution optical image texture for improved estimation of forest stand leaf area index. Canadian Journal of Remote Sensing, 22, 441±449. Wulder, M. A., Ledrew, E. F., Lavigne, M. B., and Franklin, S. E., 1998, Aerial image texture information in the estimation of deciduous and mixed forest LAI. Remote Sensing of Environment , 64, 64±76.