Single-Temporal Supervised Learning for Universal Remote Sensing Change Detection

TL;DR

This work tackles the high labeling cost of bitemporal change detection by introducing Single-Temporal Supervised Learning (STAR), which learns from unpaired single-temporal images to produce a generalizable change detector. The authors formalize pseudo-bitemporal pairs, multi-task supervision, and temporal symmetry constraints, then present ChangeStar2, a unified detector that reuses semantic segmentation architectures and incorporates temporal symmetry through a Siamese dense feature extractor and ChangeMixin2 (with temporal swap and temporal difference networks). Extensive experiments across object and semantic change detection, in-domain and cross-domain scenarios, demonstrate state-of-the-art performance on eight public datasets and show that STAR can approach or surpass bitemporal supervision in many settings, while significantly reducing labeling requirements. The work highlights STAR’s potential as a scalable, label-efficient baseline for diverse remote sensing change-detection tasks, though it acknowledges remaining challenges in out-of-domain generalization and calls for further benchmark and data-collection strategies to advance this paradigm.

Abstract

Bitemporal supervised learning paradigm always dominates remote sensing change detection using numerous labeled bitemporal image pairs, especially for high spatial resolution (HSR) remote sensing imagery. However, it is very expensive and labor-intensive to label change regions in large-scale bitemporal HSR remote sensing image pairs. In this paper, we propose single-temporal supervised learning (STAR) for universal remote sensing change detection from a new perspective of exploiting changes between unpaired images as supervisory signals. STAR enables us to train a high-accuracy change detector only using unpaired labeled images and can generalize to real-world bitemporal image pairs. To demonstrate the flexibility and scalability of STAR, we design a simple yet unified change detector, termed ChangeStar2, capable of addressing binary change detection, object change detection, and semantic change detection in one architecture. ChangeStar2 achieves state-of-the-art performances on eight public remote sensing change detection datasets, covering above two supervised settings, multiple change types, multiple scenarios. The code is available at https://github.com/Z-Zheng/pytorch-change-models.

Figures (12)

• Figure 1: Comparison of conventional bitemporal supervised learning and the proposed single-temporal supervised learning for object change detection. By exploiting object changes in arbitrary image pairs as the supervisory signals, STAR makes it possible to learn a change detector from unpaired single-temporal images.
• Figure 2: Bitemporal Image Space. Real-world bitemporal image space $\Omega(\mathbf{I}_{t+1}|\mathbf{I}_t)\times \Omega(\mathbf{I}_t)$ is a subset of the entire bitemporal image space $\Omega(\mathbf{I}_{t+1}, \mathbf{I}_t)$ because real-world bitemporal image pairs are conditioned by consistent spatial position, i.e., position consistency condition.
• Figure 3: Training sample of bitemporal supervised object change detection. (a) the image at time $t$. (b) the image at time $t+1$. (c) change label representing the change happened the time period from $t$ to $t+1$. The image $\mathbf{I}_t$ should be spatially aligned with the image $\mathbf{I}_{t+1}$.
• Figure 4: Overview of single-temporal supervised learning. The network architecture of ChangeStar2 is made up of a dense feature extractor, a semantic classifier, and ChangeMixin. ChangeStar can be end-to-end trained by segmentation loss and symmetry loss with only single-temporal supervision. The modules with the same color share the weights.
• Figure 5: Pseudo Bitemporal Image Pair. There are three typical cases of the pseudo bitemporal image pair, i.e., object change detection (1st row), semantic change detection (2nd row), and generic self-contrast (3rd row). $\mathbf{I}_t$, $\pi\mathbf{I}_t$ are the original image sequence and the new image sequence generated by a random permutation $\pi$. For change label, change regions are white and unchanged regions are black.