.png){kind=icon}
.png){kind=icon}
Return
Cross-Cycle Structured Graph Autoencoder for Unsupervised Cross-Sensor Image Change Detection
.png){kind=icon}
PRE
AI
DOI：10.1109/TAES.2026.3657315.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Unsupervised cross-sensor change detection (CSCD) is a significant yet challenging task in remote sensing, primarily due to substantial domain shifts across heterogeneous images and the difficulty of accurately modeling semantic changes. Existing methods typically rely on image translation or invariant feature extraction, where changes are indirectly inferred as reconstruction residuals. These approaches often suffer from fragile domain adaptation mappings and the entanglement of transformation errors with true changes, thereby degrading final detection performance. To address these issues, this article proposes a cross-cycle structured graph autoencoder (CC-SGAE) framework. Our model utilizes a bidirectional, cycle-consistent graph autoencoder architecture to explicitly disentangle the CSCD task from the complex domain transformation. Crucially, it introduces learnable structural difference graphs designed to directly represent semantic changes in the latent space, independent of modality-specific characteristics. The entire framework is guided by a comprehensive multicomponent loss function that enforces critical priors, including cycle-consistency, structural regularization, change sparsity, bilateral change alignment, and spatial smoothness. This explicit modeling strategy suppresses the interference of heterogeneous discrepancies and maintains the structural integrity of the detection results. Extensive experiments on five benchmark cross-sensor datasets demonstrate that our proposed CC-SGAE outperforms state-of-the-art methods, confirming its effectiveness and high potential for practical applications in unsupervised CSCD.
Keywords:
Cross-sensor change detection (CSCD)
cycle-consistency
graph autoencoder
unsupervised learning
Journal
IF:
5.7
Papers:
686
Citations:
2.4W