Assessing uncertainty in fuzzy land cover classification by confusion index

In recent years, uncertainty has become an important subject in assessing the quality of remote sensing image classification. Classification uncertainty is due to poor class definition, transition zones and the presence of mixed pixels in remote sensing data. Fuzzy classification approaches aim to estimate the proportions of specific classes that occur within each pixel. Partial class membership values derived from fuzzy classification serve as baseline information to assess classification uncertainties and allow the depiction of spatial variation of uncertainty. Providing uncertainty information at pixel level may assist in increasing the confidence in using thematic maps produced from remote sensing image classification. Many metrics have been developed to quantify pixel-wise classification uncertainty. In the present study, two formulations of confusion index are used. Literature state that, the two forms of confusion index provide similar information. The present study aims at examining whether these two formulations provide similar information or not. Multispectral image from Landsat-7 ETM + sensor was subjected to fuzzy c-means classification. The derived class membership values for each pixel were used in quantifying classification uncertainty. A comparative analysis of classification uncertainty provided by two forms of confusion index was carried out. The results from the study show that the two forms of confusion index provide dissimilar information on classification uncertainty.