A Novel Transformer-Based Multiscale Siamese Framework for High-Resolution Remote Sensing Change Detection

Change detection (CD) in remote sensing based on deep convolutional neural networks (CNNs) and transformers has played a crucial role in surface monitoring and resource development. However, the inadequate extraction of local and global features, along with the insufficient utilization of multiscale features, constrains the ability to identify change regions at multiple scales. To resolve the aforementioned issue, we propose a transformer-based multiscale Siamese framework (TMSF) for high-resolution remote sensing CD, aiming to improve detection accuracy. First, we propose a dual-attention token-based (DAT-based) transformer module, which embeds attention tokens into the transformer, effectively integrating local and global features and thereby enhancing the discriminative power of the extracted features. Second, we design a multiscale feature adaptive fusion module (MFAFM) that incorporates a channel attention mechanism to adaptively integrate multiscale information, enabling more precise identification of change regions at various scales. We verify the effectiveness of the proposed method in improving CD accuracy on three widely used benchmark datasets, LEVIR-CD, WHU-CD and CDD. Experimental results show that our method outperforms ten state-of-the-art CD methods in terms of F1-score and intersection over union (IoU) metrics. Notably, compared to that of five multiscale-based methods, our proposed TMSF achieves superior performance while requiring only half the number of parameters and computational cost. Thus, the proposed model demonstrates a marked advancement in remote sensing CD.