BiFA: Remote Sensing Image Change Detection With Bitemporal Feature Alignment

Despite the success of deep learning-based change detection (CD) methods, their existing insufficiency in temporal (channel and spatial) and multiscale alignment has rendered them insufficient capability in mitigating external factors (illumination changes and perspective differences) arising from different imaging conditions during CD. In this article, a bitemporal feature alignment (BiFA) model is proposed to produce a precise CD map in a lightweight manner by reducing the impact of irrelevant factors. Specifically, for the temporal alignment, the bitemporal interaction (BI) module is proposed to realize the alignment of the bitemporal image channel level. Our intuition is introducing the BI in the feature extraction stage may benefit suppressing the interference, such as illumination changes. Simultaneously, the alignment module based on differential flow field (ADFF) is proposed to explicitly estimate the offset of the bitemporal image and realize their spatial level alignment to mitigate the inadequate registration resulting from different perspectives. Furthermore, for the multiscale alignment, we introduce the implicit neural alignment decoder (IND) to produce more refined prediction maps achieving precise alignment of multiscale features by learning continuous image representations in coordinate space. Our BiFA outperforms other state-of-the-art methods on six datasets (such as the F1-score (F1)/intersection over union (IoU) scores are improved by 2.70%/3.91% and 2.01%/2.94% on Learning, VIsion, and Remote sensing (LEVIR)+-CD and Sun Yat-sen University (SYSU)-CD, respectively) and displays greater robustness in cross-resolution CD. Our code is available at https://github.com/zmoka-zht/BiFA.