AFPF-Net: Adjacent-Level Feature Progressive Fusion Full Convolutional Network for Remote Sensing Change Detection

In recent years, convolutional neural networks have achieved good results in the field of change detection (CD) owing to their exceptional feature extraction capabilities. However, accurately detecting objects with completely changing details, given the complex imaging conditions of bitemporal images, remains a formidable challenge. Aiming at the above challenge, we have designed a new method for remote sensing image CD. First, to capture the fine difference features at different scales, the feature difference enhancement module is proposed to enhance the information interactions not only among the bitemporal features but also between the difference features of the previous layer and the rough difference map of the current layer. Second, to accurately capture the entire region of change, the adjacent-level feature progressive fusion module is proposed, which extracts complementary information by progressively fusing high-level and low-level features, therefore enhancing the change features. Finally, based on the above two modules, a full convolution-based adjacent-level feature progressive fusion network (AFPF-Net) is designed. To validate the effectiveness of AFPF-Net, experimental evaluations are performed on two different datasets, the LEVIR-CD and WHU-CD datasets. Compared to the sub-optimal network in the experiments, the F1-score on these two datasets improved by 0.33% and 1.74%, and total model complexity is relatively reduced, achieving better balance between model performance and complexity compared to the experimental state-of-the-art network.