TANet: A Multi-Representational Attention Approach for Change Detection in Very High-Resolution Remote Sensing Imagery

The rapid progress of deep learning techniques has significantly enhanced various remote sensing tasks. However, detecting changes in very high-resolution (VHR) remote sensing images continues to be a challenge due to the complexity of the image details and limited amount of training samples. Currently, there is considerable research to achieve precise change detection (CD). This is particularly true for addressing issues such as improved detection of small targets, undulating borders and minimized noise in the output change features. To tackle these challenges, we introduce the triple-attention network (TANet), which incorporates three robust attention mechanisms; namely, channel, spatial, and frequency attentions, across both spatial and frequency domains. This is achieved by using dynamic channel attention (DCA), spatial support attention (SSA) and frequency refinement attention (FRA). In addition, we propose an advanced fusion process namely geographic matching module (GeoMM) that efficiently combines complementary information from the input bi-temporal features enabling high accuracy predictions. By integrating these powerful modules, TANet improves the model’s capacity to accurately represent changes and dynamically refine deep features at each layer by concentrating on the most critical information. Experiments on two prominent CD datasets (LEVIR-CD and WHU-CD) demonstrate that the proposed TANet outperforms previous notable approaches, with extensive ablation studies evaluating the effects of attention modules and frequency components optimization. These findings highlight TANet’s potential as a state-of-the-art solution for precise and reliable CD in VHR remote sensing imagery.