Iterative structure transformation and conditional random field based method for unsupervised multimodal change detection

• A structure transformation is proposed to transform the heterogeneous images to the same differential domain. • A CRF model is designed for multimodal change detection by incorporating the change information based unary potential, local spatially-adjacent neighbor information and global spectrally-similar neighbor information based pairwise potentials. • An iterative framework is used to combine the structure transformation and CRF segmentation to improve the accuracy. Change detection between heterogeneous images has become an increasingly interesting research topic in remote sensing. The different appearances and statistics of heterogeneous images bring great challenges to this task. In this paper, we propose an unsupervised iterative structure transformation and conditional random field (IST-CRF) based multimodal change detection (MCD) method, combining an imaging modality-invariant based structure transformation method with a random filed framework specifically designed for MCD, to acquire an optimal change map within a global probabilistic model. IST-CRF first constructs graphs to represent the structures of the images, and transforms the heterogeneous images to the same differential domain by using graph based forward and backward structure transformations. Then, the change vectors are calculated to distinguish the changed and unchanged areas. Finally, in order to classify the change vectors and compute the binary change map, a CRF model is designed to fully explore the spectral-spatial information, which incorporates the change information, local spatially-adjacent neighbor information, and global spectrally-similar neighbor information with a random field framework. As the changed samples will influence the structure transformation and reduce the quality of change vectors, we use an iterative framework to propagate the CRF segmentation results back to the structure transformation process that removes the changed samples, and thus improve the accuracy of change detection. Experiments conducted on different real data sets show the effectiveness of IST-CRF. Source code of the proposed method will be made available at https://github.com/yulisun/IST-CRF .