Land cover mapping is perhaps the most important application of remote sensing data. The abundance of mixed pixels (representing uncertainties in class allocation), particularly in coarse spatial resolution images, has always been known to lead to difficulties in producing accurate land cover classifications. Soft classification methods may help in quantifying uncertainties in areas of transition between various types of land cover. This study aims to estimate and accommodate uncertainties in all stages of a supervised classification process (i.e. training, allocation and testing) so as to produce accurate and meaningful land cover classifications. Three soft classification methods have been used-a probabilistic maximum likelihood classifier and the two classifiers based on fuzzy set theory (fuzzy c-means and possibilistic c-means). Uncertainty and accuracy measures based on a fuzzy error matrix have been adopted to evaluate each classifier. All of the classifiers show an increase in classification accuracy when uncertainty is appropriately accounted for in all stages of the supervised classification. In particular, the possibilistic c-means classifier produced the highest map and individual class accuracy and has been found to be more robust to the existence of uncertainties in the datasets. The approach suggested in this paper can be used to generate accurate land cover maps, even in the presence of uncertainties in the form of mixed pixels in remote sensing images.
Mendeley saves you time finding and organizing research
Choose a citation style from the tabs below