Urban Classification Using Full Spectral Information of Landsat ETM Imagery in Marion County, Indiana

This paper compares different image processing routines to identify suitable remote sensing variables for urban classification in the Marion County, Indiana, USA, using a Landsat 7 Enhanced Thematic Mapper Plus (ETM ) image. The ETM multispectral, panchromatic, and thermal images are used. Incorporation of spectral signature, texture, and surface temperature is examined, as well as data fusion techniques for combining a higher spatial resolution image with lower spatial resolution multispectral images. Results indicate that incorporation of texture from lower spatial resolution images or of a temperature image cannot improve classification accuracies. However, incorporation of textures derived from a higher spatial resolution panchromatic image improves the classification accuracy. In particular, use of data fusion result and texture image yields the best classification accuracy with an overall accuracy of 78 percent and a kappa index of 0.73 for eleven land use and land cover classes. Introduction Accurate image classification results are a prerequisite for many environmental and socioeconomic applications (Jensen and Cowen, 1999), such as urban change detection (Ward et al., 2000), urban heat islands (Lo et al., 1997; Weng, 2001), and estimation of biophysical, demographic, and socioeconomic variables (Lo, 1995; Thomson and Hardin, 2000). However, generating a satisfactory classification image from remotely sensed data is not a straightforward task. Many factors contribute to this difficulty, including (a) the characteristics of a study area, (b) availability of suitable remotely sensed data, ancillary and ground reference data, (c) proper use of variables and classification algorithms, (d) the analyst’s experience, and (e) the time constraint. Urban landscapes are typically composed of features that are smaller than the spatial resolution of the sensors: a complex combination of buildings, roads, grass, trees, soil, and water (Jensen, 2000). This characteristic of urban landscapes makes mixed pixels a common problem in medium spatial resolution data (between 10 to 100 m spatial resolutions) such as Landsat TM/ETM . Such a mixture becomes especially prevalent in residential areas where buildings, trees, Urban Classification Using Full Spectral Information of Landsat ETM Imagery in Marion County, Indiana Dengsheng Lu and Qihao Weng lawns, concrete, and asphalt can all occur within a pixel. The low accuracy of land cover classification in urban areas is largely attributed to the mixed pixel problem. As fine spatial resolution data (better than 5 m spatial resolution), such as Ikonos (Sugumaran et al., 2002; van der Sande et al., 2003) and ADAR digital multispectral scanner imagery (Thomas et al., 2003), become easily available in recent years, they have been increasingly utilized for urban studies. A major advantage of these fine spatial resolution imageries is that such data greatly reduce the mixed pixel problem, providing the potential to extract much more detailed information of urban structures compared with medium spatial resolution data. However, a new severe problem comes with the fine spatial resolution image data, i.e., the shade problem caused by topography, tall buildings, or trees (Asner and Warner, 2003). Moreover, high spectral variation within any land-cover class often decreases the classification accuracy. The huge amount of data storage and severe shade problem in fine spatial resolution image give rise to challenges for selection of suitable image processing approaches and classification algorithms over a large area. Last but not least, high spatial resolution imagery is much more expensive and requires much more time to implement data analysis than medium spatial resolution image data. In practice, medium spatial resolution imagery, especially the TM/ETM being readily available for multiple dates, is still the most commonly used data for urban classification, in spite of the mixed pixel problem. Different approaches have been used in order to improve urban classification accuracy. These approaches can be roughly grouped into four categories: (a) use of sub-pixel information (Rashed et al., 2001; Phinn et al., 2002), (b) data integration of different sensors or sources (Harris and Ventura, 1995; Haack et al., 2002), (c) making full use of the spectral information of a single sensor (Gong and Howarth, 1992; Stuckens et al., 2000; Shaban and Dikshit, 2001), and (d) use of expert knowledge (Stefanov et al., 2001; Hung and Ridd, 2002). Traditional per-pixel classification algorithms, such as maximum likelihood classifier (MLC), are still frequently used for image classification due to their simplicity and availability. However, much attention has been shifted to develop more advanced techniques and classification algorithms, such as neural network, contextual, object oriented, and knowledge-based classification approaches in PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Novembe r 2005 1275 Dengsheng Lu is with the Center for the Study of Institutions, Population, and Environmental Change, Indiana University, Bloomington, Indiana 47408 ([email protected]). Qihao Weng is with the Department of Geography, Geology, and Anthropology, Indiana State University, Terre Haute, Indiana 47809 ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 71, No. 11, November 2005, pp. 1275–1284. 0099-1112/05/7111–1275/$3.00/0 © 2005 American Society for Photogrammetry and Remote Sensing 04-006.qxd 10/11/05 1:31 AM Page 1275