Heterogeneous Change Detection with Self-Supervised Deep Canonically Correlated Autoencoders

This paper proposes a new method for bitemporal change detection in heterogeneous remote sensing images. A modified canonical correlation analysis is used to align the code layers of two deep convolutional autoencoders, one for each image domain. It weights the input with a new affinity-based prior, which measures changes in pixel relations across the image domains and is used to reduce the influence of data points prone to change. By this procedure of self-supervision, we adapt the intrinsically supervised architecture to the unsupervised case, noting that the censoring of change pixels is key to efficiently learning the required data transformations. The result is an unsupervised algorithm which allows change detection in either of the image domains, or a combination of those, since efficient domain translation is obtained by coupling cross-domain encoders and decoders. We demonstrate state-of-the-art performance on real test datasets.