HIFANet: A Hierarchical Feature Aggregate Network for Change Detection of High-Resolution Remote Sensing Images

Change detection (CD) is a crucial task in remote sensing, requiring high precision and accuracy. Traditional CNN-based siamese network have been successfully applied to CD tasks in numerous research studies. However, these methods still face several challenges. Firstly, mainstream CD networks lack effective image feature extraction capabilities, despite the rich object details present in high-resolution remote sensing images. Secondly, as the networks become deeper, there are issues with the loss of image details and the increasing number of network parameters, which impact training costs. To address these challenges, we propose a hybrid architecture CD network called Hierarchical Feature Aggregate Network (HI-FANet). Through the integration of the Convolution-Involution Module (CIM) and Transformer structures, our approach not only enhances the capability of extracting informative features from images but also addresses the limitation of inadequate feature extraction in existing methods. Moreover, compared to a purely Transformer-based architecture, our hybrid network achieves comparable performance while significantly reducing the number of parameters, resulting in a substantial reduction in the training cost. In addition, our network employs skip-dense connections to facilitate feature fusion, demonstrating superior capability in mitigating the loss of fine-grained image details compared to alternative fusion methods. Furthermore, we introduce the Efficient Attention Predict Head (EAPH) by combining Channel Attention Module (CAM) in a residual way to address semantic differences between features, resulting in better detection results. Experimental results on the CDD and LEVIR datasets demonstrate that our proposed method achieves higher accuracy and F1 scores compared to existing methods.