Multi-Objective Evolutionary Ensemble Clustering using Improved Log-ratio Operator for SAR Images Change Detection

The speckle noise existing in synthetic aperture radar (SAR) image will have a strong impact on the accuracy of change detection. Thus, a multi-objective evolutionary ensemble clustering method based on an improved log-ratio operator is proposed to improve accuracy of SAR image change detection. First of all, to suppress speckle noise and gain as much image information as possible from the original image pair, an improved log-ratio operator is designed by taking neighborhood pixels information into account to generate the difference image. Secondly, SAR image change detection is considered as a multi-objective optimization problem due to the paradox between intra-class compactness and inter-class separability from the point of clustering. One of objective functions, using local prior probability and local neighborhood entropy, is constructed to raise detection performance. In addition, according to the designed selection solution index, the clustering results corresponding to Pareto optimal solutions are selectively integrated by incorporating locally weighted graph partitioning (LWGP) model. Finally, a specially designed neighborhood error correction strategy is applied in the ensemble result to acquire the ultimate change detection result. The experimental results on three real SAR datasets confirm that the proposed method has a more effective and superior performance compared with the other popular algorithms.