An Unsupervised Change Detection Approach Based on KI Dual Thresholds under the Generalized Gauss Model Assumption in SAR Images

The unsupervised change detection technique on multi-temporal SAR images not only needs to detect the changed region but also subdivide the changed region in a complex geographical environment so that the backscatter enhanced class and the backscatter weakened class can be further identified.The generalized Gaussian distribution model can approximate a large variety of statistical distributions at the cost of only one additional parameter to be estimated(i.e.,the shape parameter) compared with the traditional Gaussian distribution model.In particular,the generalized Gaussian distribution model is proved to be more suitable to describe the distributions of unchanged and changed classes on SAR log-ration difference image than the Gaussian one.An change detection algorithm in SAR images based on the generalized Gaussian distribution model and KI dual thresholds criterion is proposed.The probability density distributions of the unchanged class,the backscatter enhanced class and the backscatter weakened class on SAR difference image are modeled.The dual thresholds criterion function is defined based on KI criterion.An optimal automatic dual thresholds selection approach is proposed only using the gray histogram of the difference image.The unchanged,the backscatter enhanced and the backscatter weakened classes are detected.The two temporal SAR images from Radarsat satellite are used to experiment and the results show that the proposed approach is feasible and effective.Improving the accuracy and speed of SAR image unsupervised change detection technique by using the spatial context information will be studied as a future development of this work.