REMOTE SENSING IMAGES CHANGE DETECTION USING THE SIAMESE NETWORK COMBINED WITH PURE SWIN TRANSFORMER

While Transformers have become prevalent in detecting changes in remote sensing imagery, several challenges hinder their broader adoption in the field, such as missing detection, low precision of contour boundary detection and the complexity of calculation in the data processing. Addressing these challenges, our research introduces the SiamFormer architecture, which employs a layered Swin Transformer encoder coupled with a cascaded decoder for feature fusion, specifically designed for change detection within remote sensing imagery. First, our approach integrates the Siamese network framework with a pure Swin Transformer, crafting a decoder with a hierarchical layout to bolster its capabilities for pixel-level change detection in remote sensing images. Second, we improve the interconnection status between decoder layers by top-down cascade paths and dense cascades to produce high-quality high-resolution image semantic feature change outputs. Furthermore, to mitigate the loss of edge details in change objects caused by high-dimensional downsampling, we implement a convolutional decoding classifier. This classifier efficiently reduces the channel dimensions of the merged change feature map to the bare minimum. Our experimental analysis, conducted on the CDD and LEVIR-CD datasets, demonstrates that our proposed methodology outperforms existing change detection techniques for remote sensing imagery in terms of effectiveness. © 2024, Politechnica University of Bucharest. All rights reserved.