An Efficient GPU-Parallelized Network for Bi-Temporal Remote Sensing Change Detection via Visual State Space Models

High-resolution remote sensing change detection (RSCD) is a core task for applications such as urban planning and natural disaster monitoring. Although deep learning has achieved significant progress on multi-temporal data, existing architectures still face severe challenges. Convolutional neural networks (CNNs) are limited by their local receptive fields and thus struggle to capture long-range spatial dependencies; Transformer-based models offer global modeling ability but their self-attention incurs quadratic spatiotemporal complexity O≤ft(L2), which causes prohibitively large GPU memory consumption when processing large, high-resolution images. Traditional recurrent models (RNNs/LSTMs) can model temporal dynamics but their sequential computation prevents effective use of modern GPUs' parallelism. To address these issues, we propose a dual-temporal change detection network for bi-temporal remote sensing images, named VMamba-CD, built upon a visual state-space model (VMamba). VMamba employs a visual scanning mechanism to perform temporal enhancement and global feature extraction on each input timestamp independently, achieving coverage of a global receptive field while keeping computation complexity linear, O(L). After dual-path feature extraction, a differential fusion module aggregates change cues, and a lightweight convolutional upsampling decoder restores spatial details for precise change localization. Experimental results show that VMamba-CD outperforms mainstream models on multiple metrics, achieving an Intersection-over-Union (IoU) of 89.40%, an F1 score of 94.34%, and an overall accuracy (OA) of 99.08 %. The proposed method not only resolves the computational bottleneck of Transformers but also overcomes CNNs' limited receptive fields and the poor parallelizability of LSTMs, offering a hardware-friendly, efficient alternative for high-resolution remote sensing change detection.