Texture-Integrated Classification of Urban Treed Areas in High-Resolution Color-Infrared Imagery

lkaditional multispectral classification methods have not provided satisfying results for treed area extraction from highresolution digital imagery because trees are characterized not only by their spectral but also by their textural properties. Treed areas in urban regions are especially dificult to extract due to their small area. Many other urban objects, such as lawn and playgrounds, cause confusion because they display similar, even identical, spectral properties. In this study a texture integrated classification method is proposed. To effectively extract tree textural features and eliminate noise, an algorithm of conditional variance detection is developed, which consists of a directional variance detection and a local variance detection. This algorithm detects tree features with higher accuracy than common texture algorithms. By integrating the new algorithm with traditional multispectral classification, treed areas in urban regions can be extracted with sufficiently high accuracy. Application of the new approach in different urban areas indicates that the average accuracy of treed area extraction was increased from 67 percent, using a multispectral classification, to 96 percent, using the texture integrated classification. Introduction Inventory and mapping of urban treed areas is important for urban environment study and urban planning. Traditionally, urban treed areas are extracted and mapped through visual interpretation of aerial color-infrared images and fieldwork (Nowak et al., 1996). Multispectral classification methods have been used to extract urban treed areas from digitized aerial images, but successful results have rarely been achieved because of the variety of spectral and textural properties of trees (Hildebrandt, 1996). Many studies on classification of medium-resolution satellite imagery (e.g., Landsat TM, SPOT) in urban or urban-rural areas demonstrated that treed areas, except large wooded areas within big parks and suburban areas, could not be extracted because of the lack of spatial resolution (Barnsley and Ban; 1996; Gao and Skillcorn, 1998). In some studies, urban treed areas were extracted by including texture information. The extracted treed areas could, however, not be separated from lawn areas (Gong and Howarth, 1990; Zhang, 1998a). Numerous high-resolution airborne sensors have been developed. The world's first commercial high-resolution satellite imagery, IKONOS, has been available since the fall of 1999. However, the improvement of the spatial resolution does not automatically increase classification accuracy when conventional multispectral classification methods are used (Marceau et al., 1994), because high spatial resolution increases spectral and spatial variation within individual classes. By analyzing Department of Geodesy and Geomatics Engineering, University of New Brunswick, P.O. Box 4400, Fredericton, New Brunswick E3B 5A3, Canada ([email protected]). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING intraclass local variances in images with different resolutions, Woodcock and Strahler (1987) suggested that, in urban and forested environments, techniques utilizing texture modelling were appropriate for images with a resolution of 10 to 30 m; Marceau et al. (1994) concluded that an average 10-m spatial resolution was outimal for suectral classification of a tem~erate forested envGonment, whle higher spatial resolutioh'did decrease the classification accuracy. To classify images in urban environments with a higher iesolution (ag., leis than 5 m), new algorithms and tools considering textural information are consequently required. Such algorithms and tools are especially important for dealing with the problems arising from the rapid emergence of high-resolution imagery. Many studies on extracting detailed tree information from high-resolution real or simulated images have utilized tree textural features in the extraction. For example, neural network (Dreyer, 1993), co-occurrence matrix (Anys et al., 1994), valley following (Gougeon, 1995), semivariogram (St-Onge and Cavayas, 1997), threshold-based spatial clustering (Culvenor et al., 1999), local maximum filtering (Wulder et al., 2000), and local variance (Coops and Culvenor, 2000) were employed to identify tree textures. However, reliable texture extraction or successful tree recognition in real remote sensing data is difficult (Bruniquel-Pine1 and Gastellu-Etchegorry, 1998; Pinz, 1999). Although some approaches worked well with high-resolution images containing near fully forested areas, the success rate reduced rapidly in the areas where man-made structures are located (St-Onge and Cavayas, 1997), especially in urban areas (Anys et al., 1994). To separate buildings from treed areas, digital surface models (DSMS) generated by airborne laser scanners were employed as main or auxiliary sources in some studies. For example, Hahn and Statter (1998) extracted treed areas according to both the spectral information in airborne color-infrared images (0.5m resolution) and the height information in laser DSMs. B r u ~ and Weidner (1997) used the geometric information of trees in a DSM to discriminate trees from buildings. Haala et al. (1998) extracted buildings and treed areas using laser DSMs and maps. However, airborne laser DSMs are currently expensive and not available for many cities. Using geometric information in DSMs, it is difficult to separate trees from buildings when they are close to each other. In addition, information on the location of buildings available in a map is not always current. In this study, a simple but effective method is proposed to extract urban treed areas directly from high-resolution colorinfrared images without using any auxiliary height information. This method integrates the textural and spectral information of Photogrammetric Engineering & Remote Sensing Vol. 67, No. 12, December 2001, pp. 1359-1365. 0099-1112/01/6712-1359$3.00/0 O 2001 American Society for Photogrammetry