MGLFA-Net: multi-scale global–local feature aggregation network for remote sensing change detection

Change detection (CD) in remote sensing (RS) imagery is a critical task that has seen significant advancements with the evolution of deep neural networks. However, current methods that combine convolutional neural networks and transformer modules often fall short in achieving high CD performance, particularly in detecting small changes, due to insufficient multi-scale information. Furthermore, although transformer-based models are effective at capturing global context, they come with high computational costs. To overcome these challenges, we propose a multi-scale global–local feature aggregation network, which includes a multi-scale Siamese encoder, a feature difference enhancement fusion module (FDEFM), and a decoder. This architecture leverages comprehensive contextual information to precisely identify change regions while minimizing the size of the model. Our encoder design integrates global context and local spatial details using deformable convolution and simplified self-attention mechanisms. In addition, the FDEFM effectively captures and emphasizes correlated changes between image pairs, enhancing CD sensitivity and noise suppression, thus improving the model’s performance. Experimental results on three challenging CD datasets demonstrate that our approach achieves state-of-the-art performance.