GNLF-Net guided non-local feature enhancement network for remote sensing change detection

In recent years, convolutional neural networks (CNNs) have made significant progress in the field of remote sensing image change detection (CD) due to their outstanding feature extraction capabilities. However, detecting subtle changes in the detailed regions of bi-temporal remote sensing images, especially under complex imaging conditions, remains a significant challenge. To address this challenge, we propose a new remote sensing image change detection network—GNLF-Net (Guided Non-Local Feature Enhancement Network). GNLF-Net is based on a U-Net framework, with the encoder built on the VGG16 structure, extracting multi-scale features of image pairs in parallel. The core design aims to enhance change detection accuracy by extracting multi-scale features and guiding the modeling of global information while focusing on key regions. First, to capture the subtle differences between bi-temporal images, we introduce the Guided Non-Local Gaussian Attention module. This module dynamically adjusts the weight of non-local information modeling through guidance and then smooths the features, enabling the model to focus more on the change areas. During the decoding process, we adopt a step-by-step feature fusion strategy, integrating multi-scale information through a series of decoding modules. Finally, we generate accurate change detection results through a high-resolution reconstruction process. To validate the effectiveness of GNLF-Net, we conducted experimental evaluations on two publicly available remote sensing image change detection datasets—LEVIR-CD and WHU-CD. The experimental results show that GNLF-Net performs excellently in change detection tasks, achieving an IoU score improvement of 0.48% and 1.21%, respectively, compared to existing SOTA methods. The code is available at: https://github.com/wxy115646/GNLF-Netl.