Efficient multiscale feature integration network for lightweight remote sensing images change detection

Remote sensing change detection (RSCD) identifies surface changes between bi-temporal images, but existing methods face high computational costs that hinder real-world deployment. We propose EMFINet, an efficient lightweight network with four novel modules: (1) Multiscale Feature Extraction Module (MFEM) extracts discriminative multiscale features via parallel convolutions with adaptive dilation rates; (2) Difference Feature Enhancement Module (DFEM) enhances change-related signals and suppresses noise through dual-attention mechanisms; (3) FFT-aided Global Context Perception Module (FGCPM) models long-range dependencies by fusing spatial self-attention with frequency-domain learning via FFT; (4) Adaptive Decoding Module (ADM) refines change maps through semantic-guided adaptive feature fusion. Extensive experiments on LEVIR-CD, WHU-CD, and CDD datasets show EMFINet outperforms fifteen state-of-the-art methods, achieving F1 scores of 91.66%, 94.56%, and 96.74% with only 2.44 M parameters and 3.69 G FLOPs. The source code is available at https://github.com/gcm139/EMFINet.