An unsupervised approach based on Riemannian metric to change detection on multi-temporal SAR images

This paper investigates the problem of detecting changes on multitemporal SAR imagery in an unsupervised way. A novel change indicator was developed to identify the temporal changes. It is computed by the local average of the amplitude ratio comparing the exponentiation of the local average of the logarithm-transformed amplitude ratio. Compared with the classical ratio of local means, the novel operator is more effective in identifying the changed pixels even the local means are preserved. The classification is implemented by an automatic thresholding algorithm derived from a new Riemannian metric defined in the differential geometry structure. The geodesic distance derived from the new Riemannian metric provides a way to compare the distance between the probability distributions of the changed class and the non-changed class. The probability density functions of the changed and non-changed classes are characterized over the photometric variable. By maximizing the distance between the probability density distributions of the two classes, the misclassification errors are minimized and the optimal threshold is achieved accordingly. Experiments were carried on portions of multi-temporal Radarsat-1 SAR data. The obtained accuracies confirm the effectiveness of the proposed approach.