A NOVEL GRAPH BASED FUZZY CLUSTERING TECHNIQUE FOR UNSUPERVISED CLASSIFICATION OF REMOTE SENSING IMAGES

This paper addresses the problem of unsupervised land-cover classification of multi-spectral remotely sensed images in the context of self-learning by exploring different graph based clustering techniques hierarchically. The only assumption used here is that the number of land-cover classes is known a priori. Object based image analysis paradigm which processes a given image at different levels, has emerged as a popular alternative to the pixel based approaches for remote sensing image segmentation considering the high spatial resolution of the images. A graph based fuzzy clustering technique is proposed here to obtain a better merging of an initially over-segmented image in the spectral domain compared to conventional clustering techniques. Instead of using Euclidean distance measure, the cumulative graph edge weight is used to find the distance between a pair of points to better cope with the topology of the feature space. In order to handle uncertainty in assigning class labels to pixels, which is not always a crisp allocation for remote sensing data, fuzzy set theoretic technique is incorporated to the graph based clustering. Minimum Spanning Tree (MST) based clustering technique is used to over-segment the image at the first level. Furthermore, considering that the spectral signature of different land-cover classes may overlap significantly, a self-learning based Maximum Likelihood (ML) classifier coupled with the Expectation Maximization (EM) based iterative unsupervised parameter retraining scheme is used to generate the final land-cover classification map. Results on two medium resolution images establish the superior performance of the proposed technique in comparison to the traditional fuzzy c-means clustering technique. 1