MNDAF-Net: Multiscale Neighborhood Decoupling and Attribute-Focused Fusion for Dual-Temporal Remote Sensing Image Scene Change Detection Network

Existing methods in remote sensing image scene change detection typically rely on convolutional or Transformer architectures to extract attribute features, combining local details with global semantic modeling. However, these approaches often overlook the importance of neighborhood attribute features for capturing local changes and contextual stability, leading to insufficient representation of fine-grained changes. In addition, bitemporal feature fusion commonly suffers from information loss, limiting the depth of interaction modeling in change regions and impacting both accuracy and robustness. To address these challenges, we propose a novel dual-temporal remote sensing scene change understanding network with multiscale neighborhood decoupling and attribute-focused fusion. In the encoding phase, we leverage a multiscale neighborhood feature decoupling analysis (NFDA) module to enhance independent modeling of neighborhood attributes, effectively capturing fine-grained changes in complex scenes. Specifically, this module integrates multihead adjacent atrous attention (MA2) and a panoramic adaptive attribute sensing (PA2S) layer to provide comprehensive multiscale perception and adaptive feature enhancement for bitemporal remote sensing images. In the decoding phase, the attribute-focused fusion decoder (AFF-Decoder) enhances the fusion of bitemporal features by ensuring correlated interactions between dual-temporal features, capturing fine-grained relationships to achieve more accurate change representations. Experimental results on three public datasets—LEVIR-CD, WHU-CD, and DSIFN-CD—demonstrate that the proposed MNDAF-Net achieves promising performance in remote sensing attribute change understanding. Our code will be published at https://github.com/lzp-lkd/MNDAF-Net