DIVISION S-5—PEDOLOGY Delineation of Soil Variability Using Geostatistics and Fuzzy Clustering Analyses of Hyperspectral Data

for this. First of all, it is difficult to obtain images at the optimal time for delineating bare soil patterns, and there A soil map is one of the key data layers for developing a robust is always some degree of mixture effects with vegetation global model and evaluating land quality and use. A current soil map produced by conventional soil survey is the major source of soil and dry plant components. Second, the spatial and specinformation. However, such a map may not provide the desired accutral resolution of the existing systems are often too racy in terms of scale and cartographic quality as a digital format coarse to delineate soil boundaries at the desired detail. for geographic information system (GIS) modeling applications. This Lastly, there are always some discrepancies between study was designed to introduce and test the procedures for improving the map derived from remote sensing data analysis and the objectivity and accuracy in the delineation of soil patterns with the existing soil maps because the map derivation of the use of hyperspectral imagery. These hyperspectral data were analyzed two systems is completely different for classifying soils. through different models including the linear mixture model, blockThe current hierarchical classification system of soils kriging interpolation, and fuzzy-c-means (FCM) algorithms. Hyperbased on soil taxonomy is not natural in that the forming spectral remote sensing data, having very good spectral and spatial processes of a soil are not the same as those of biological resolution, were used for quantifying soil patterns and conditions. A linear spectral mixing model was effectively used not only for reducing evolution. Soil boundaries and polygons are not discrete dimensionality but also for removing vegetation effects for studying and homogeneous but continuous and heterogeneous soil patterns from a single soil map layer derived from hyperspectral (Burrough, 1989; Odeh et al., 1992a,b). remote sensing data. Block kriging interpolation based on a semivarioThe hyperspectral imaging spectrometer may provide gram fitted with the isotropic exponential model represented soil more detailed information of specific soil properties, patterns very well beyond the limitation of the size of pixel. Fuzzyenabling us to create detailed soil spectral delineations. c-means clustering analysis showed clear membership patterns and Mixed pixel problems can be reduced by using the linear segmented soil patterns effectively, although this is not a soil map in spectral mixture model. A soil component map layer the conventional sense. without vegetation effects could be used for further spatial analysis. Advanced geostatistical methods involving analysis of the semi-variogram and kriging have R soil information is a critical factor for debeen applied to soils with considerable success (Burgess veloping a robust global model and evaluating land and Webster, 1980a,b; Burgess et al., 1981; Di et al., use more efficiently. Furthermore, an accurate soil map 1989; Odeh et al., 1992a,b; Webster and Oliver, 1990). is one of the key data layers in implementing these Kriging enables us to create a detailed spectral soil map kinds of analyses with GIS tools (Fisher, 1987). Aerial beyond the limitation of sizes of pixels by interpolating photography has been used as useful ancillary data for unsampled locations. Fuzzy-c-means clustering analysis delineating soil boundaries. Often today’s soil informawill be used for classifying the final kriged spectral pattion requirements cannot be met with existing soil maps tern since the concept of fuzzy theory is useful for dealbecause of the lack of detailed soil information that can ing with the delineation of ambiguous soil polygon be derived from these maps. In addition, the advent of boundaries (Burrough, 1989; Burrough et al., 1992; precision farming requires a more accurate and objecOdeh et al., 1992a,b). This research is designed to create tive method for improving and supplementing existing a detailed representation of soil patterns with hypersoil maps (Indorante et al., 1996). spectral remote sensing data through kriging interpolaRemote sensing has often been used for delineating tion and fuzzy classification. soil boundaries and polygons. The applicability of this technology has not made significant advances but it may MATERIALS AND METHODS be valuable for providing information about timedependent aspects of the soil such as moisture status Study Area and Data Source and erosion (Burrough, 1991). There are several reasons The parent material of the nearly level upland soils of the study area was glacial till. These soils are deeply leached and C.-W. Ahn, University Space Research Association/NASA Goddard were formed under prairie vegetation. The spectral data were Space Flight Center, Biospheric Sciences Branch Code 923, Greenbelt, obtained on 12 June 1992 over Tippecanoe County, Indiana, MD 20771; M.F. Baumgardner, Dep. of Agronomy, Purdue Univ., with NASA’s Airborne Visible/Infrared Imaging SpectromeWest Lafayette, IN 47907-1150; L.L. Biehl, School of Electrical & Computer Engineering, Purdue Univ., West Lafayette, IN 47907. This research was supported by NASA Research Grant NAGW-3862. ReAbbreviations: AVIRIS, Airborne Visible/Infrared Imaging Specceived 16 Jan. 1997. *Corresponding author ([email protected]). trometer; DN, digital number; FCM, fuzzy-c-means; GIS, geographic information system; GPS, Global Positioning System. Published in Soil Sci. Soc. Am. J. 63:142–150 (1999).