Full fuzzy land cover mapping using remote sensing data based on fuzzy c-means and density estimation

Abstract

The three stages in supervised digital classification of remote sensing data are training, classification, and testing. The commonly adopted approaches assume that boundaries between classes are crisp and hard classification is applied. In the real world, however, as spatial resolution decreases significantly, the proportion of mixed pixels increases. This leads to vagueness or fuzziness in the data, and in such situations researchers have applied the fuzzy approach at the classification stage. Some researchers have tried fuzzy approaches at the training, classification, and testing stages (full fuzzy concept) using statistical and artificial neural network methods. In this paper a full fuzzy concept has been presented, at a subpixel level, using density estimation using support vector machine (D-SVM) and fuzzy c-means (FCM) approaches. These approaches (SVM and FCM) were evaluated with respect to a fuzzy weighted matrix. In this test study using a four-channel dataset, a comparison of methods has found that a D-SVM function using a Euclidean norm yields the best accuracy.