Opportunities of Using Random Sets to Model Uncertainties in Agricultural Field Boundaries Observed from Remote Sensing Images

Random sets are common spatial statistical concepts that allow quantifying uncertainty in spatial objects. For objects extracted from remote sensing images, quantification of the uncertainty is important, as many objects are relatively small with respect to the pixel size and are sometimes poorly defined. Remote Sensing (RS) data are important in land cover identification, classification and estimation. The aim of this paper is to address problems associated with the presence of edges between objects. Such edges occur on images in different shapes, for example as borders between agricultural parcels. The study was applied on an NDVI map of a Landsat 5 TM image. Field boundaries are normally irregular and often transitional. Modeling agricultural fields as spatial objects helps to identify the extensional uncertainties and therefore to characterize inaccuracy in parcel size estimation. The study was carried out in the Sharifabad region in Iran. The Douglas Paucker algorithm was used to establish a single boundary that separates different parcels of agricultural fields. The results of the study indicate that Gaussian thresholding of image segmentation generated random sets for six agricultural fields. Quantification of extensional uncertainty presented two parcels with a larger extensional uncertainty than the other four parcels. A question we addressed in this study was identification of the boundaries between two adjacent parcels. An overall accuracy of 91% shows that random sets were effective for modeling the extensional uncertainty of the agricultural fields and for the delineation of the agricultural field boundaries. We conclude that the geometric model used to delineate the agricultural field boundaries is able to properly handle irregular shape boundaries.