Urban Land-use Classification Using Variogram-based Analysis with an Aerial Photograph

In this study, a variogram-based texture analysis was tested for classifying detailed urban land-use classes, such as mobile home, single-family house, multi-family house, industrial, and commercial from a digital color infrared aerial photograph. Spectral classification was first carried out to separate the building class from non-building classes. Then, a building-presence binary image was generated so that building pixels were assigned a value of “1” and non-building pixels were assigned a value of “0.” Multiple texture bands were further generated employing a variogram-based texture analysis and used for land-use classification. The generation of the building presence binary image allowed us not only to fully explore the capability of variogram-based analysis on spatial pattern detection, but also to prevent the variogram-based analysis from being disturbed by the natural fluctuation of spectral signals. The result from using a mosaic test image was considered satisfactory with a kappa coefficient of 0.72. Introduction The urban environment contains a variety of spectrally different materials, such as soil, grass, trees, plastic, metal, shingle wood, and concrete. Traditional pixel-based spectral classifications assign each pixel to one of the candidate classes based on its brightness value, which indicates the spectral reflectance of the earth surface. However, using spectral information alone is not sufficient for classification of spectrally heterogeneous land-use classes, such as mobile home, single-family house, multi-family house, industrial, and commercial. Therefore, texture information from spatial patterns is often used to complement spectral information. New texture bands, in addition to original spectral bands, may be used together in classification. Each pixel in each of the texture bands is assigned a digital value as a reflection of the spatial variation of pixel brightness in the neighborhood in a sense of describing local textures. There are several approaches of creating texture bands. The variogram of geostatistics, applied in a window-based Urban Land-use Classification Using Variogram-based Analysis with an Aerial Photograph Shuo-sheng Wu, Bing Xu, and Le Wang function, is a relatively recent technique (Miranda et al., 1992; Miranda and Carr, 1994). The variogram is commonly represented by a graph of semi-variance against the lag. The lag is the distance between paired data points. The semi-variance is half the average of the squared difference between paired data values. The mathematical function of the semi-variance ( ) by certain lag (h) can be expressed as (Burrough and McDonnell, 1998): where n is the number of paired pixels; z (xi) and z (xi h) are pixel values at xi and xi h, respectively. Variograms allow remote sensing researchers to measure the degree of spatial autocorrelation inherent in different landscapes as recorded in remote sensing images. However, the spectral values calculated in the variogram may be influenced by multiple factors, such as soil type, solar radiation, precipitation run off, flooding frequency, and wind direction. Each factor may have its own sub-variogram representing a unique autocorrelation structure, while a combined effect will create a summed, observed variogram. Burrough and McDonnell (1998) thus suggested that a pragmatic approach was to use a domain-specific variogram whenever possible. For this reason, we generated a buildingpresence variogram to describe spatial patterns alone. Rather than computing variograms directly using pixel gray values, we computed the variogram based on a binary image of buildings. The binary image will replace the original spectral image to serve as a base to extract autocorrelation information contributed by the building pattern alone. The rationale of using this binary image was not only to remove the effect of spectral fluctuation caused by a number of factors, but also to highlight and reserve the spatial patterns of buildings. There have been two main approaches of using the variogram as texture measures in the field of remote sensing. One approach was that the variogram was modeled by a mathematical function and the coefficients of the function were used as texture measures. Some example studies included Ramstein and Raffy (1989), Herzfeld and Higginson (1996), Chen and Stow (2002), and Chen and Gong (2004). The other approach was to use semi-variance values at various lags as texture measures. Some example studies (h) 1 2na n i 1 cz(xi) z(xi h) d 2 PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J u l y 2006 813 Shuo-sheng Wu and Le Wang are with the Department of Geography, Texas State University-San Marcos, 601 University Drive, San Marcos, TX 78666-4616 (sw1020@ txstate.edu; [email protected]). Bing Xu is with the Department of Geography, University of Utah, 260 S. Central Campus Dr., Rm. 270, Salt Lake City, UT 84112-9155 and also with the Department of Environmental Science and Engineering, Tsinghua University, Beijing, 100084, China ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 72, No. 7, July 2006, pp. 813–822. 0099-1112/06/7207–0813/$3.00/0 © 2006 American Society for Photogrammetry and Remote Sensing 04-103 6/9/06 9:55 AM Page 813