Study On Hyperspectral Remote Sensing Application In Recognizing Urban Material

Diverse spectra of material in urban is more complicated than natural environment's, especially for various decades, difference of materials and its composition. Conventional multispectral data associated with statistical data-driven methods are ineffective for extracting urban land cover information in detail due to the insufficiency on spectral and spatial resolution. Through spectral analysis and spectral matching technology supported by the geographical information, land cover and artificial targets of urban can be classified discriminatingly. Meanwhile, space information could be more efficient to help hyperspectral data interpret the real material of terra by correlation among various matters. Results of classification provide information about such as city planning, environmental monitor, and variance of city.As the first spaceborne hyperspectral sensor, NASA's EO-1 Hyperion now provide images in 242 spectral bands in the 400-2500 nm range compared to multispectral sensors such as Landsat ETM+ which provide only 6 broadbands over this wavelength region. The research in this study obtains the reflectance of land cover material from a series of preprocessing, geometric correction, radiance calibration, band selection, and atmospheric correction via FLAASH which can remove 'smile' effect. Spectra of field measurements can hardly apply in classification directly because of main affection by several geographical aspects, such as topography, climate. Particularly, existing of mixed pixels results in the less widespread of field measurements. On the other hand, end members in image can easily classify or identify the hyperspectral data.To regard as precondition of understanding hyperspectral data, end member extraction is always the hotspot. Methods already exist mainly aim at semi-auto or auto extract end members via non-empirical. In this research, a pixel-based spectral analysis algorithm, spectral angle mapper (SAM), is adopted to extract end members by matching image spectra to experimental spectra (by John Hopkins university spectral library). Before application of SAM, image should be masked by water and vegetation to eliminate their affection in follow-up procedure. Density slice is applied in the rule images which is half-results of the SAM, in order to detect the end members in least angle. Simultaneously, extract the same endmembers through purest pixel index (PPI), a kind of statistical method. Choose six pixels from every end member calculated from SAM and PPI for precision examining based on hyper spatial image data. Then locate six end members, validate them by means of comparison of hyper spatial image data, QuickBird, photographed at the near time. It is found that this method is more exactly to extract end members, and makes the endmembers match the real material more reasonably. In the end, abundance of surface material in Guangzhou city is decomposed by linear spectral mixture model (LSMM) approach, and surface material map is derived from setting range for abundance map. Because of static formula set in the urban land cover, DEM calculated from the ASTER stereo pairs shall be easily used to analyze the mixed spectral pixel as the tool of auxiliary. Then recalculate the abundance map to produce classification map through programming it in IDL language. The results confirm the ability of hyperspectral in urban material mapping, and can clear classify the material in urban with the total precision 76.2099%, Kappa efficiency is 0.7258, as follow: cement concrete, paving concrete, clay tile roof, outdated building, bare soil, high reflectivity material(e.g. glass, new metal), low reflectivity material(e.g. shadow), forest, grass, and water. Otherwise, supported by spatial information hyperspectral data can exert its further advantages. Making use of the Gram-Schmidt sharpening fusion method between hyperspectral and hyperspatial data can not only increase the spatial resolution of hyperspectral data, but also maintain the integrity of spectral feature preferably. Moreover, the method founded on DEM calculate the classification more precisely than without any ancillary spatial data. The precision of classification is improved up to 82.4268%, and the Kappa efficiency is as high as 0.7833.