Shannon entropy based fuzzy distance norm for pixel classification in remote sensing imagery

Pixel classification of mixed pixels in overlapping regions of remote sensing images is a very challenging task. Efficiency and detection of uncertainty are always the key ingredients for this task. This paper proposes an approach for pixel classification using Shannon's entropy-based fuzzy distance norm. Unsupervised clustering is used to group the objects based on some similarity or dissimilarity. The proposed algorithm is able to identify clusters comparing fuzzy membership values based on Shannon's entropy evaluation. This new normalized definition of the distance also satisfies separability, symmetric and triangular inequality conditions for a distance metric. This approach addresses the overlapping regions in remote sensing images by uncertainties using fuzzy set membership values. Shannon entropy further introduces belongingness and non-belongingness to one cluster within the distance measure. We demonstrate our algorithm for segmenting a LANDSAT image of Shanghai. The newly developed algorithm is compared with FCM and K-Means algorithms. The new algorithm generated clustered regions are verified with on hand ground truth facts. The validity and statistical analysis are carried out to demonstrate the superior performance of our new algorithms with K-Means and FCM algorithms.