A NEW UNSUPERVISED CHANGE DETECTION APPROACH WITH HYBRID CLUSTERING FOR DETECTING THE AREAL DAMAGE AFTER NATURAL DISASTER

摘要

In this study, an effective unsupervised change detection approach, which use optical images, based on combined difference images and hybrid clustering, for detection the areal damage of natural disasters such as flooding and fire, has been proposed. First, absolute difference operator and absolute logarithmic difference operator is applied to co-registered images to compose two kinds of change maps. Secondly, a feature space is created from two kinds of change maps with a combination function. Then, median filter and wiener filter are applied to feature space, respectively. The median filter is used to maintain the edge information and the wiener filter is used to eliminate isolated pixel and considering local consistency in order to detect areal damage more accurately. After that, min-max normalization is applied to filtered data to increase the effectiveness of Backtracking Search Optimization Algorithm (BSA). Lastly, filtered data is clustered with hybrid approach into two classes, damaged and undamaged area. In clustering stage, firstly, temporary cluster centroids are calculated with K-Means. Then, optimal cluster centroids are obtained with BSA through a cost function. Local consistency and edge information of the combined image are considered in this approach. Experimental result obtained from two real optical data sets prove the effectiveness of the proposed approach.