Mid and Low Resolution SAR Image Change Detection Based on Fractal and Multi-scale Analysis

For pixel based SAR image change detection, the discrepancy images produced by the log ratio operation or Kullback-Leibler divergence cannot achieve satisfactory results in artificial target change detection. We introduce a fractal dimension into the construction of discrepancy images and define the Fractal Dimension-Log Ratio (FD-LR) image capable of detecting changes both from the natural targets and the artificial targets. A Gaussian mixture distribution is used to model the statistical properties of FD-LR. The Bayesian principle with expectation maximization-based parameter estimation is conducted to perform unsupervised thresholding on the FD-LR. To reduce speckle interferences , multiscale analysis and data fusion in the decision step are performed. Comparative experiments confirm the effectiveness of the proposed approach.