Difference Enhancement and Dependence-Aware Network for Remote Sensing Image Change Detection

In the context of the booming development of new remote sensing sensor technologies, change detection (CD) using bitemporal remote sensing images has an undeniable role in monitoring changes of surface objects. However, the current deep-learning-dominated methods have not efficiently extracted and utilized the difference features between bitemporal images, leading to the loss of important foreground information and redundancy in background information. In addition, classic challenges such as pseudochange and spectral confusion in complex scenes also weaken the performance of CD. This article proposes a difference enhancement and dependency-aware network (DEDANet) for remote sensing image CD. The network introduces a bitemporal difference aggregator that highlights regions of variation while filtering out noise interference to obtain a more complete characterization of the differences. To focus on foreground information, multiple foreground information enhancers are inserted after the siamese feature extractor in the network to perform multiscale feature enhancement before extracting difference features. To deepen the perception of complex dependencies, we further design a dual-domain context fusion module, which models the temporal dependencies between bitemporal images based on a visual state-space model and introduces frequency-domain convolutions to supplement information in gradient areas. DEDANet significantly outperforms the existing optimal algorithms as evaluated on three standard datasets.