Consistency Change Detection Framework for Unsupervised Remote Sensing Change Detection
Yating Liu, Yan Lu
TL;DR
The paper tackles unsupervised change detection in bi-temporal remote sensing imagery, where labels are unavailable and generator-based reconstruction can memorize input-output mappings. It introduces the Consistency Change Detection Framework (CCDF), combining a Cycle Consistency (CC) module with bidirectional generators and a Semantic Consistency (SC) module to preserve reconstruction details, along with a Change Segmentation (CS) network to identify unreconstructable regions. Training proceeds in three stages: Stage 1 performs global style transfer with cycle-consistent generators, Stage 2 jointly trains CS and SC to detect changes while preserving semantic detail, and Stage 3 alternates fine-tuning of the generators and segmentation model; losses include $L_{gen}$, $L_{cyc}$, and $L_{cont}$ via a pre-trained network, plus $L_{seg}$, $L_{reg}$, and $L_{sem}$. On WH and HY GF-2 datasets, CCDF achieves state-of-the-art F1 and cIOU scores, with ablation studies confirming that CC reduces generator overfitting and SC improves detail coherence, highlighting the practical value of label-free, robust change mapping for urban monitoring and disaster response.
Abstract
Unsupervised remote sensing change detection aims to monitor and analyze changes from multi-temporal remote sensing images in the same geometric region at different times, without the need for labeled training data. Previous unsupervised methods attempt to achieve style transfer across multi-temporal remote sensing images through reconstruction by a generator network, and then capture the unreconstructable areas as the changed regions. However, it often leads to poor performance due to generator overfitting. In this paper, we propose a novel Consistency Change Detection Framework (CCDF) to address this challenge. Specifically, we introduce a Cycle Consistency (CC) module to reduce the overfitting issues in the generator-based reconstruction. Additionally, we propose a Semantic Consistency (SC) module to enable detail reconstruction. Extensive experiments demonstrate that our method outperforms other state-of-the-art approaches.