Change Detection on Remote Sensing Images Using Dual-Branch Multilevel Intertemporal Network

Multilevel feature aggregation and enhancement network for remote sensing change detection

“…To evaluate the effectiveness and efficiency of our MFAENet, we select several state-of-the-art (SOTA) models as competitors, including CNNs-based methods: FC-EF, 19 FC-SD, 19 FC-SC, 19 SNUNet, 49 TFI-GR, 45 Changer, 37 MFIN, 56 MFSFNet, 57 CICNet, 58 and two transformer-based methods: BIT 31 and ChangeFormer 33 . We reproduce these models using their publicly available code and default parameters to ensure fair comparisons.…”

“…However, none of the above methods establishes the long-range dependency in the features, which in turn affects the detection performance. Feng et al 50 unified self-attention and cross-attention in a single module, proposed a cross-temporal joint attention block. This block serves as a guide for the global feature distribution of each input.…”

Remote-Sensing Image Change Detection Based on Adjacent-Level Feature Fusion and Dense Skip Connections

“…Meanwhile, in order to enhance the network's ability to extract change features, many improvement methods have been proposed on how to fuse the dual-temporal multi-scale features obtained after the Siamese network, including feature exchange [33], differential enhancement [34], [35], and various types of attention mechanisms, such as self-attention [36], the combination of self-attention [9] and cross-attention [9], spatial attention [37], and the combination of spatial attention and channel attention [38], [39], and selective attention [40], etc. For example, ECFNet [35] adjusts the channels of all features to a uniform value after the difference operation of dual-temporal multi-scale features and then fuses them from bottom to top so that the features of different scales contribute the same to the final result during the fusion process.…”

“…STANet [9] introduces a self-attention module for calculating positional relations between any two pixels in dual-temporal features and uses them to generate the change maps at different scales. DMINet [36] unifies self-attention and cross-attention in the JoinAtt module, which focuses the attention on the change region and suppresses the noise independent of the change. Zhao et al [41] proposed a three-branch network TSNet, which uses a Siamese network and a single-stream network to respectively extract the features of dual-temporal images and their spliced images and uses a dual-channel attention module to fuse the three types of features to extract change features.…”

Impact of user-generated travel posts on travel decisions: A comparative study on Weibo and Xiaohongshu

“…Recent advancements in deep learning have precipitated a shift towards network-based change detection methods. Leveraging the capabilities of neural networks, these methods are capable of extracting robust image features and performing spatiotemporal fusion across various feature extraction stages effectively [13,33]. Early deep learning-based algorithms perceived change detection as a dense prediction task, with the majority exploring different stages of bi-temporal feature fusion to generate the final change segmentation map.…”