UCDFormer: Unsupervised Change Detection Using Real-time Transformers

Change detection (CD) by comparing two bitemporal images is one of the most challenging tasks in remote sensing. With the advantages of requiring no cumbersome labeled change information, unsupervised CD has attracted extensive attention from researchers. However, rarely existing unsupervised CD approaches consider the seasonal differences in multi-temporal images. To this end, we propose a change detection with a domain shift setting for remote sensing images. Furthermore, we propose a novel unsupervised CD method using a real-time transformer, called UCDFormer. Specifically, a transformer-driven image translation is first proposed to mitigate domain shift between two images with real-time efficiency. Importantly, to effectively reduce the translation strength of changed pixels and to increase the translation of unchanged pixels, a domain-specific affinity weight is computed in an unsupervised fashion for a weighted translation loss. After image translation, we can generate the difference map between the translated before-event image and the original after-event image. Then, classification maps from the difference map are obtained by fuzzy c-means (FCM) clustering and adaptive threshold. However, these detection results are not very reliable. Thus, a novel reliable pixel extraction module is proposed to select significantly changed/unchanged pixel positions by fusing the results of FCM clustering and adaptive threshold. These selected pixel pairs are used to learn a binary classifier, and the other pixel pairs will be classified by this classifier to obtain the final CD results. Experimental results on different unsupervised CD tasks with seasonal changes demonstrate the effectiveness of the proposed UCDFormer compared with several other related methods.