AnyTime-CD: Self-supervised change detection in arbitrary-length dense time series of high-resolution remote sensing images

Abstract

With the continuous enhancement of the observation capabilities of remote sensing satellites, dense multi-temporal high-resolution remote sensing images have become an important data source for monitoring surface changes. However, multi-temporal change detection (MTCD) remains challenging due to inconsistent time intervals, spatial misalignments, and a scarcity of labeled data, all of which hinder the generalization and practical deployment of deep learning models in large-scale applications. To address the above challenges, this paper proposed a multi-task self-supervised change detection framework (i.e., AnyTime-CD), which supports the input of time series of any length and can jointly model temporal-spectral-spatial features without manual annotation. This framework consists of three sub-tasks: temporal dynamic modeling, spectral semantic consistency learning, and spatial feature alignment, which respectively address the issues of irregular time interval, spectral perturbation, and spatial offset, thereby collaboratively optimize network performance. We evaluated AnyTime-CD on two globally distributed multi-temporal high-resolution datasets. The experimental results showed that, under an unsupervised setting, this method achieved performance close to or even surpassing that of the supervised methods. Compared with the best-performing self-supervised method, it achieved relative gains of over 13% (spatial) and 8% (temporal) in F1 scores (using a one-frame tolerance strategy), respectively. After sample fine-tuning, it further achieved F1 scores of 86.40% (spatial) and 79.50% (temporal), exceeding the state-of-the-art supervised method by 4.44% and 9.09%, respectively. Furthermore, AnyTime-CD showed good adaptability to different pre-training configurations. It is worth noting that it exhibited strong robustness under medium and low cloud cover conditions and could even utilize cloud perturbations as data augmentation to further enhance performance. In conclusion, AnyTime-CD offers a flexible and label-free solution for MTCD tasks, suitable for complex, dynamic, and interference-prone remote sensing scenarios, and is expected to promote the application of self-supervised methods in surface monitoring.