Unsupervised Heterogeneous Change Detection in Radar Images by Cross-domain Affinity Matching

A new methodology for unsupervised heterogeneous change detection has recently been proposed, which combines deep neural networks for domain alignment and image-to-image regression with a comparison of domain-specific pixel affinities to reveal structural changes. In this paper we explain the underlying cross-domain dissimilarity measure and give an example of how it can be integrated in a deep neural network architecture and used to strengthen approaches based on image transformations. We demonstrate that the method is viable for bitemporal change detection with synthetic aperture radar (SAR) images combined with optical images, as well as with combinations of multifrequency and multipolarisation SAR images, but also provide a critical assessment of weaknesses and remaining challenges.