Remote Sensing Images Change Detection Based on Level Set Model

Threshold methods are commonly used in traditional unsupervised change detection. Favorable change detection results can be obtained in general. However, this method is only applicable to the situations where the changed and the unchanged areas have high contrast. When the contrast is low, the change detection results can be seriously affected. The change detection task is formulated as a segmentation issue where the discrimination between the changed and unchanged classes is achieved by defining an energy function. The minimization of the function is carried out by using a level set method to find a global optimal contour, which can split the image into two mutual exclusive regions associated with changed and unchanged classes respectively. The complete energy function of the LCVLS (A Variational Level Set Model Based on Local Clustering) is composed by energy items using global clustering criterion, curve length, regularization item and the penalty function. Experimental results show that the LCVLS model is more effective than other unsupervised change detection methods.