Hierarchical Feature Alignment-based Progressive Addition Network for Multimodal Change Detection

Abstract

Multimodal Change Detection (MCD) has become a hot topic and thus enhanced much attention. Different from homogeneous change detection, MCD needs to identify changes by comparing heterogeneous Bi-Temporal Images (BTIs) of different modalities. Although some patch-level MCD methods have been reported, there are still few studies on large-scale image-level MCD methods. The main challenge of MCD is that it is more difficult to capture the differences between different modality images than heterogeneous BTIs. To address the challenge, this paper proposes a novel non-Siamese Hierarchical Feature Alignment-based Progressive Addition Network (HFA-PANet) for MCD. In the proposed HFA-PANet, two novel modules are devised to elevate the difference features of multimodal BTIs, thereby improving its change extraction capability. First, a Hierarchical Feature Alignment Module (HFAM) based on multiple kernel maximum mean discrepancy is integrated into each level to reduce domain shift, which achieves feature alignment of heterogeneous BTIs and obtains the difference features through the aligned features. Then, we devise a Progressive Addition Module (PAM) to gradually aggregate the difference features at each level to enhance the comparability between changed areas and backgrounds. In addition, we combine hierarchical domain alignment loss and hybrid multi-level loss to train the proposed model, which effectively improves the MCD performance of the proposed method. Extensive experiments on two publicly available large-scale MCD datasets show that the proposed HFA-PANet can achieve performance gains compared with other State-Of-The-Art (SOTA) and popular approaches. The source code of the proposed HFA-PANet will be available after publication at https://github.com/TongfeiLiu/HFA-PANet-for-MCD.