A Dual-Stream UNet With Parallel Channel–Spatial Interaction and Aggregation for Change Detection

Change detection (CD) aims to analyze changes in objects between bitemporal images, playing a crucial role in geographical observations. However, hidden difference features and insufficient inter-feature interaction both lead to suboptimal performance of methods. To address these challenges, this paper proposes a dual-stream UNet with parallel channel-spatial interaction and aggregation (DCSI-UNet). Firstly, the encoding part utilizes a multilayer dual-stream convolutional neural network (CNN) to extract multilayer features of bitemporal images and then processes them through a channel group interaction module (CGIM) and a spatial Gaussian attention module (SGAM) to generate channel and spatial interaction feature maps, respectively. Specifically, the CGIM splits the bitemporal feature maps into multiple channel-wise subspaces and applies interactive attention to enhance cross-channel feature correlation. The SGAM calculates the mixed mean and joint variance at the spatial level of bitemporal feature maps, which establish a Gaussian attention kernel capturing hidden difference features. Secondly, the decoding part is designed in two phases: skip connections and interaction-feature aggregation. The former decodes the interactive feature maps from CGIM and SGAM separately. The latter fuses multilayer channel and spatial interactive features from the encoding part, which suppresses localization loss and exploits richer semantic consistency of changed regions. Finally, comprehensive experiments on three public datasets demonstrate that the proposed method outperforms existing methods in both quantitative and qualitative evaluations. The code is available at https://github.com/ZChaoyv/DCSI-UNet.