Land-use/Land-cover Classification with Multispectral and Hyperspectral EO-1 Data

We compared the capability of the Earth Observing-1 (EO-1) Hyperion hyperspectral (HS) data with that of the EO-1 Advanced Land Imager (ALI) multispectral (MS) data for discriminating different land-use and land-cover classes in Fremont, California. We designed a classification scheme of two levels with level I including general classes and level II including more specific classes. Classification shows that the HS data does not produce better results than the MS data when we directly applied a Mahalanobis distance (MD) classifier. We tested a number of feature reduction and extraction algorithms for the HS image. These algorithms include principal component analysis (PCA), segmented PCA (SEGPCA), linear discriminant analysis (LDA), segmented LDA (SEGLDA), penalized discriminant analysis (PDA) and segmented PDA (SEGPDA). Feature reductions were all followed by an MD classifier for image classification. With SEGPDA, SEGLDA, PDA, and LDA, similar accuracies were achieved while a segmentation-based approach we proposed (SEGPDA or SEGLDA) greatly improved computation efficiency. They all outperformed SEGPCA and PCA by 4 to 5 percent (level II) and 1 to 3 percent (level I) in classification accuracy. For level II classification, overall accuracies obtained by using the features extracted from the HS image were 2 to 3 percent greater than those obtained with the MS image. For various vegetation class and impervious land use categories, the HS data consistently produced better results than the MS data. For level I classification, the HS image generated a thematic map that is 0.01 greater in kappa coefficient comparing to the MS image. When we collapsed the level II classification map to a level I map, 5 percent (HS) to 7 percent (MS) improvements were achieved. Introduction The EO-1 Hyperion is the only hyperspectral (HS) sensor operated in space (NASA, 1996). An HS sensor has contiguous narrow wavelength bands (about 10 nm each) that are able to capture more subtle spectral details of the objects on the ground than a multispectral (MS) sensor (about 100 nm Land-use/Land-cover Classification with Multispectral and Hyperspectral EO-1 Data Bing Xu and Peng Gong each). However, it is often difficult to process the high dimension of such data ( 200 bands). The limited number of training samples comparing to the high dimension of data and the high correlation among the adjacent bands will lead to inaccurate estimation of the covariance structures and degenerate ranks of spectral matrices, thus limiting the accuracy of classification (Hughes, 1968; Hsu et al., 2002). Land-use classification in an urban setting using multispectral data or panchromatic imagery has been widely studied. Spectral, contextual, texture, and structural information are extracted to aid the characterization of different and complex land surfaces and to improve the accuracy of identification (Gong and Howarth, 1990a and 1990b; Gong and Howarth, 1992; Deguchi and Sugio, 1994; Ridd, 1995; Xu et al., 2003). Gong et al. (1997, 2001, and 2002) measured in situ HS data in order to establish a spectral library, to recognize different species of conifers, and to further extract ecological parameters by data transformation and selection of biophysically sensitive bands. Thenkabail et al. (2004) reported substantial improvement in rainforest type classification in Africa when using Hyperion data compared to the use of Advanced Land Imagery (ALI), Ikonos, and Landsat ETM data. However, analysis of imaging spectrometer data for urban land-use applications has rarely been explored. Roessner et al. (2001) applied the concept of linear spectral unmixing, considering spatial neighborhood using a procedure of iterative endmember selection to differentiate land surfaces in an urban area in Germany. Therefore, two questions lead us to this research. First, does the HS image contain more information than the MS image for urban land-use classification? This question has not yet been addressed probably due to the unavailability of simultaneous acquisition of both the MS and HS data. The launch of EO-1 made it possible for us to explore and compare the classification results from the two sensors. Second, as for HS data, what feature reduction/extraction methods would be better in terms of both computational efficiency and classification accuracy? Direct spectral matching using binary coding and vectorization with known spectra in a spectral library to label the unknown pixels is possible because of the high spectral resolution (Jia, 1996; Goetz et al., 1985). Due to the complexities caused by the high dimensional space of the data, feature extraction schemes such as PCA or fisher’s LDA have been applied in transforming and reducing the data dimension by maximizing the ordered variance of the whole PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Augu s t 2007 955 Bing Xu is with the Center for Natural and Technological Hazards, Department of Geography, University of Utah, 260 S. Central Campus Dr., Rm. 270, Salt Lake City, UT 841129155 ([email protected]). Peng Gong is with the Department of Environmental Science, Policy, and Management, 137 Hilgard Hall, University of California, Berkeley, CA 94720 and the State Key Lab of Remote Sensing Science, Jointly Sponsored by Institute of Remote Sensing Applications, Chinese Academy of Sciences, and Beijing Normal University, Beijing, 100101, P.R. China ([email protected]). Photogrammetric Engineering & Remote Sensing Vol. 73, No. 8, August 2007, pp. 955–965. 0099-1112/07/7308–0955/$3.00/0 © 2007 American Society for Photogrammetry and Remote Sensing 04-098.qxd 7/10/07 3:07 PM Page 955