Using Spatial Co-Occurrence Texture to Increase Forest Structure and Species Composition Classification Accuracy

should be augmented with texture measures (St-Onge and The analysis of forest structure and species composition with Cavayas, 1995). Texture can be used in the classification or, high spatial resolution (5 1 m) multispeckal digital imagery alternatively, the process of per-pixel classification can be supis described in an experiment using spatial co-occurrence plemented or replaced with another approach, such as texture texture and maximum-likelihood classification. The segmentation (Lobe, 1997). Much work remains to be done on objective was to determine if higher forest species composition how and where texture analysis can be effective in f o r e s ~ classification accuracies would result in comparison to the remote sensing applications (Lark, 1996; ~ u l d e r et al., 1996; use of spectral response patterns alone. Increased accuracy Bmiquel-Pinel and Gastellu-EtchegO~* lgg8). was obtained when using texture at all levels of a classifi'cation A large degree of variability can exist in the development hierarchy. At the stand level, accuracies were on the order of of a classification signature using high spatial resolution image 75 percent in agreement with field surveys, an improvement data. In Figure 1, approximately 1-m pixels are shown of a forof 21 percent over the accuracy obtained using spectral data est stand adjacent to a logging road Tintersection. Large Stanalone; in stands grouped according to species dominancelcodard deviations, relative to the mean spectral response, typidominance, the accuracy improved still firther to 80 percent. cally result in forested scenes in most spectral bands. The diffiThe overall classification accuracy in a highly generalized culties in using spectral signatures comprised of the mean and lifeform classification was 100 percent. This represented a 33 standard deviation are obvious, giving rise to the notion that percent increase in accuracy over that which could be some measure of spatial variability would be useful in signaobtained, in a classic spectral "signature" classification ture generation (Figure 2). Texture analysis attempts to meaapproach, using spectral response patterns alone. sure this scene variance for use in the classification process. Traditionally, texture has been defined as the spatial variation Introduction in image tones or colors (Haralick et al., 1973). In images of forThe use of high spatial resolution remote sensing data to claseStcOver9 spatial variation may be causedb~ changesin species sify forest structure (St-Onge and Cavayas, 1997) and identify tYPel crown stem density. forest species composition (Franklin et al., 2000) within forest Different stem densities can create different texture patstands continues to interest forest scientists, managers, and terns* even have the same practitioners (Hudson, 1987; Sali and Wolfson, 1992; Baulies ( F i ~ e 3). In the first set (Figure 3; top row labeled as group and Pons, 1995; Leckie et al., 1995; Meyers et al., 1996; Pitt and lB), the first image contains data from a mature spruce et al., 1997). The objective of achieving accuracies that meet or stand with 500 stems per hectare. The exceed those currently achieved by aerial photointerpretation image (goup lB) data a Vruce plantation* but techniques (Congalton and Mead, 1983; Gillis and Leckie, with approximately 3000 stems per hectare. In the second set 1993) may soon be feasible given continued improvements in 3; midd1e row labeled as 2A and 2B) the difference computing, image and image analysis techbetween two hardwood stands with different crown closures is niques (Green, 2000). Image classification is one possible shown. Group 2A illustrates image data from a mature intolermethod for use in this application. One of the challenges in ant stand with a that was using high spatial resolution remotely sensed imagery in digiclosed. Group 2B was acquired over a stand which had a much t d classification is that the interclass spectral variability of SUmore open canopy* with a measured crown closure of 30 to 50 face features can increase with increasing spatial resolution. Percent. The open canopy created a larger shadow component The result is a reduction in class statistical separability (Hay thanthat re~resentedb~ the first in different et al., 1996). With traditional classifiers, which rely on the conFigure mixed-w00d cept of a signature," this often translates into a poor stands (bottom row* labeled as groups 3A and 3B). First* group classification accuracy for individual tree species (Hughes 3A is a 70 percent hardwood and 30 percent softwood mixedet al., 1986) and aggregated estimates of species composition stand. This stand can be to goup 3B9 a mixed(Franklin, 1994). In general, it is thought that, rather than relystand with 60 percent and 40 percent harding on multispectral image spectral signatures done, digital wood. Each of these stands had similar structure (i.e., shadow classification of species composition and forest structure S.E. Franklin and A. J. Maudie are with the Department of GeogPhotogrammetric Engineering & Remote Sensing raphy, University of Calgary, Calgary, Alberta T2N 1N4, Canada VO~. 67, NO. 7, July 2001, pp. 849-855. ([email protected]). 0099-lll2/01/6707-849$3.00/0 M.B. Lavigne is with the Canadian Forest Service, Atlantic 6 2001 American Society for Photogrammetry Forestry Centre, Fredericton, New Brunswick E3P 5P7, Canada. and Remote Sensing PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING luly 2001 849