31 December 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
In response to the existing challenges in semantic change detection for remote sensing images, such as weak spatiotemporal correlation and insufficient utilization of local neighborhood information, this paper proposes a semantic change detection network based on hierarchical local-sparse attention (HLSNet). The network combines a fully convolutional network (FCN) with a deep transformer structure to leverage the advantages of local feature extraction and long-range information connection. Next, a hierarchical local-sparse attention is proposed to exploit the neighborhood characteristics of target pixels using a dual-window attention mechanism, the aim is to increase the receptive field while minimizing the interference of redundant information. By focusing on all tokens within a smaller window and dynamically selecting key tokens within a larger window for attention calculation, this two-tiered attention approach allows the model to handle details while capturing broader contextual information. The small window provides tightly related local information, while the larger window offers relevant but potentially more distant information, achieving a hierarchical processing of information from local to long-range. In order to facilitate more comprehensive interaction between the features of pre- and post-change images, each transformer block in the network employs a strategy of concatenating self-attention and cross-attention. This approach better captures the spatiotemporal correlations and feature integration, thus achieving efficient and precise change detection. HLSNet achieves the highest accuracy on the two commonly used SCD datasets, SECOND and Landsat-SCD, with F scd values reaching 62.53% and 91.67%, respectively.