Predicting the Areal Extent of Land-Cover Types Using Classified Imagery and Geostatistics

Remote sensing is an efficient means of obtaining largemeans to obtain these data in a timely and consistent manner. Yet, remotely sensed land-cover data are not errorarea land-cover data. Yet, remotely sensed data are not error-free. This paper presents a geostatistical method to free, as they rely largely on the spectral responses of landcover types that may not all be spectrally distinguishable. model spatial uncertainty in estimates of the areal extent of land-cover types. The area estimates are based on exhausData accuracy may further degrade as a result of errors in the source data and imperfect image processing. If retive but uncertain (soft) remotely sensed data and a sample of reference (hard) data. The method requires a set of motely sensed land-cover data are used to evaluate environmental changes one should, therefore, account for the unmutually exclusive and exhaustive land-cover classes. Landcover regions should be larger than the pixels’ ground resocertainties in these data. Foody et al. (1992), Maselli et al. (1994), Van der lution cells. Using sequential indicator simulation, a set of equally probable maps are generated from which uncerWel et al. (1998), De Bruin and Gorte (2000), and others explored how posterior probability vectors, a by-product of tainties regarding land-cover patterns are inferred. Collocated indicator cokriging, the geostatistical estimation probabilistic image classification, can be used to represent local uncertainty about class labels of individual pixels. This method employed, explicitly accounts for the spatial crosscorrelation between hard and soft data using a simplified paper goes one step further and presents a geostatistical approach to assess spatial uncertainty (Goovaerts, 1997, model of coregionalization. The method is illustrated using a case study from southern Spain. Demonstrated uncertain1999; Deutsch and Journel, 1998), that is, the joint uncertainty about land cover at several pixels taken together. ties concern the areal extent of a contiguous olive region and the proportion of olive vegetation within large pixel This is particularly useful in regional analyses that require spatially aggregated land-cover data. Examples of these are blocks. As the image-derived olive data were not very informative, conditioning on hard data had a considerable effect assessments of the areal extent of land-cover types over spatial units with fixed geometry (e.g., political units or on the area estimates and their uncertainties. For example, the expected areal extent of the contiguous olive region square cells) or the size of contiguous regions having one vegetation cover. Sequential indicator simulation (SIS) enincreased from 65 ha to 217 ha when conditioning on the reference sample. Elsevier Science Inc., 2000 ables the generation of multiple maps that honor the available data and allow spatial patterns and uncertainties in the mapped land cover to be inferred. Because in SIS spatial structures are described in terms of variograms, the INTRODUCTION approach is notably different from the one proposed by Current concerns about environmental changes have lead Canters (1997), who used image segmentation to derive to an increased demand for land-cover data at regional to spatial structures. global scales (e.g., DeFries and Townshend, 1994; VogelRecently, Kyriakidis (1999) used SIS to map thematic mann et al., 1998). Satellite remote sensing is an efficient classification accuracy through integration of imagereported (soft) and higher accuracy (hard) class labels. Data integration was accomplished by using simple indicator * Wageningen UR, Centre for Geo-Information, Wageningen The Netherlands kriging with varying local means (SKlm) (Goovaerts and Address correspondence to S. de Bruin, Wageningen UR, Centre Journel, 1995; Goovaerts, 1997) obtained from spatially for Geo-Information, P.O. Box 47, 6700AA Wageningen, AH, Netherdegraded classified imagery. In this study, the soft indicator lands. E-mail: [email protected] Received 17 November 1999; revised 10 March 2000. data are derived from an image classifier’s posterior proba-