Unsupervised Change Detection in multitemporal Satellite Images: A VMamba-Driven Cross-Scale Feature Decoding Network

Unsupervised change detection based on deep learning has received wide attention on remote sensing image analysis. However, a critical challenge that limits the accuracy of change detection, is how to establish the global and local feature representations of multitemporal remote sensing images in complex scenarios. To address this challenge, a VMamba-driven cross-scale feature decoding (VCFD) unsupervised change detection network is proposed. VCFD employs two weight shared VMamba encoders to extract multi-scale features from multitemporal remote sensing images, which models the global context and local details of change features. Meanwhile, we design a cross-scale upsampling decoding module to progressively reconstruct high-resolution feature maps. The experiment on the OSCD dataset shows the superior performance of proposed method, which achieves a new existing state-of-the-art(SOTA) in unsupervised change detection.