An approach for mapping large-area impervious surfaces: Synergistic use of Landsat 7 ETM+ and high spatial resolution imagery

A wide range of urban ecosystem studies, including urban hydrology, urban climate, land use planning and resource management require current and accurate geospatial data of urban impervious surfaces. We developed an approach to quantify urban impervious surfaces as a continuous variable by using multi-sensor and multi-source datasets. Subpixel percent impervious surfaces at 30-meter resolution were mapped using a regression tree model. The utility, practicality and affordability of the proposed method for large-area imperviousness mapping were tested over three spatial scales (Sioux Falls, South Dakota, Richmond, Virginia, and the Chesapeake Bay areas of the United States). Average error of predicted versus actual percent impervious surface ranged from 8.8 to 11.4% with correlation coefficients from 0.82 to 0.91. The approach is being implemented to map impervious surfaces for the entire United States as one of the major components of the circa 2000 national land cover database. INTRODUCTION The status and trends of urban land cover and land use significantly impact the quality of human life and urban ecosystems. Accurate, up-to-date and spatially explicit data on urban land cover and land use are required to support urban land management decision-making, ecosystem monitoring and urban planning (Ridd, 1995). One of the most important land cover types characteristic of urban and suburban environment is the impervious surfaces developed through anthropogenic activities. Impenetrable surface, such as rooftops, roads and parking lots have been identified as a key environmental indicator of urban land use and water quality (e.g. Arnold and Gibbons,1996). The spatial extent and distribution of impervious surfaces impact urban climate by altering sensible and latent heat fluxes within the urban surface and boundary layers; Impervious surface also increases the frequency and intensity of downstream runoff and decreases water quality. Strong correlation between imperviousness of a drainage basin and the * This work was performed under U.S. Geological Survey contract 1434-CR-97-CN-40274. This paper is preliminary and has not been edited or reviewed for conformity with U.S. Geological Survey standards or nomenclature. U.S. Department of the Interior U.S. Geological Survey 2 quality of its receiving streams has been reported. For example, stream quality usually starts to degrade if more than ten percent of the area of a watershed is impervious (Schueler, 1994). In recognizing its environmental significance, impervious surface has been identified as one of the major components of the circa 2000 National Land-Cover Data base (NLCD 2000) to be developed through the Multi-Resolution Land Characteristics (MRLC) 2000 Consortium (Homer et al., 2002). The MRLC 2000 consortium was formed to meet the needs of several federal agencies of the United States (U.S. Geological Survey, Environmental Protection Agency, USDA Forest Service, NASA and NOAA) for Landsat 7 Enhanced Thematic Mapper Plus (ETM+) imagery and land cover/land use data. Through the MRLC 2000 consortium, agencies formed a partnership and pooled resources to develop: 1) a multi-temporal Landsat 7 ETM+ image dataset containing three dates of imagery per path-row for the United States, and 2) a consistently developed circa 2000 national land cover database. PREVIOUS STUDIES OF URBAN IMPERVIOUS SURFACES Numerous research efforts have been devoted to quantify urban impervious surfaces using ground-measured and remotely sensed data (Deguchi and Sugio, 1994; Williams and Norton, 2000; Phinn et al., 2000). The methodologies range from multiple regression (Foster, 1980; Ridd, 1995), spectral unmixing (Ji and Jensen, 1999; Ward et al., 2000), artificial neural network (Wang, 2000; Flanagan and Civco, 2001), classification trees (Smith and Goetz, 2001), and integration of remote sensing data with geographic information systems (Prisloe et al., 2001). Ridd (1995) proposed a conceptual model, i.e., vegetation-impervious surfacesoil (VIS) for urban ecosystem analysis. This framework presents a systematic standard for characterizing urban ecosystem from morphological, biophysical, and anthropogenic perspectives. Using this model detailed land cover land use and biophysical parameters were obtained for urban ecosystems using remote sensing data. Forster (1980) examined the relationship between Landsat MSS data and percent land cover types sampled at the pixel level from the Sydney metropolitan area using multiple regression techniques. He found that variables most closely correlated with intensity of urban developed areas were those of normalized band ratios. More recent studies adopted advanced machine learning algorithms and spectral unmixing that allow the derivation of imperviousness at the sub-pixel level. For instance, Flanagan and Civco (2001) conducted a subpixel impervious surface mapping using artificial neural network and an ERDAS Imagine subpixel classifier. For four municipal study areas in Connecticut, the overall accuracy at impervious-non-impervious detection level varied from 71 to 94% with a root mean square error (RMSE) of 0.66 to 5.97%. Wang et al. (2000) developed a U.S. Department of the Interior U.S. Geological Survey 3 subpixel proportional land cover information transformation (SPLIT) model, a modularized artificial neural network-based algorithm, to quantify proportion of land cover types from high-resolution multispectral videography. Overall accuracy achieved was 87.6%. Spectral unmixing and classification trees classifier have also been capable of quantifying sub-pixel impervious surface. The accuracy of imperviousness estimates from unmixing was also comparable (Ji and Jenson, 1999; Ward et al. 2000), whereas overall within-class accuracy using classification trees was about 84% in a study of Montgomery County, Maryland (Smith and Goetz, 2001). Thus far almost all research conducted was confined within a limited spatial area (one urban setting or at county-level), and each study used only one type of data for developing training data for model prediction.