[논문 리뷰] DiffUCD:Unsupervised Hyperspectral Image Change Detection with Semantic Correlation Diffusion Model

Xiangrong Zhang, Shunli Tian외 3명|arXiv (Cornell University)|2023. 05. 21.

Remote-Sensing Image Classification인용 수 8

한 줄 요약

DiffUCD introduces an unsupervised hyperspectral image change detection framework that uses a semantic correlation diffusion model and cross-temporal contrastive learning to achieve state-of-the-art results among unsupervised methods, approaching supervised performance.

ABSTRACT

Hyperspectral image change detection (HSI-CD) has emerged as a crucial research area in remote sensing due to its ability to detect subtle changes on the earth's surface. Recently, diffusional denoising probabilistic models (DDPM) have demonstrated remarkable performance in the generative domain. Apart from their image generation capability, the denoising process in diffusion models can comprehensively account for the semantic correlation of spectral-spatial features in HSI, resulting in the retrieval of semantically relevant features in the original image. In this work, we extend the diffusion model's application to the HSI-CD field and propose a novel unsupervised HSI-CD with semantic correlation diffusion model (DiffUCD). Specifically, the semantic correlation diffusion model (SCDM) leverages abundant unlabeled samples and fully accounts for the semantic correlation of spectral-spatial features, which mitigates pseudo change between multi-temporal images arising from inconsistent imaging conditions. Besides, objects with the same semantic concept at the same spatial location may exhibit inconsistent spectral signatures at different times, resulting in pseudo change. To address this problem, we propose a cross-temporal contrastive learning (CTCL) mechanism that aligns the spectral feature representations of unchanged samples. By doing so, the spectral difference invariant features caused by environmental changes can be obtained. Experiments conducted on three publicly available datasets demonstrate that the proposed method outperforms the other state-of-the-art unsupervised methods in terms of Overall Accuracy (OA), Kappa Coefficient (KC), and F1 scores, achieving improvements of approximately 3.95%, 8.13%, and 4.45%, respectively. Notably, our method can achieve comparable results to those fully supervised methods requiring numerous annotated samples.

연구 동기 및 목표

• Motivate unsupervised HSI-CD due to labeling cost and imaging-condition variability.
• Leverage diffusion models to capture semantic correlations in spectral–spatial features.
• Mitigate pseudo changes from imaging conditions via cross-temporal contrastive learning.
• Fuse semantic-spatial features with spectral-difference-invariant features for accurate CD.
• Demonstrate state-of-the-art performance on multiple public datasets without labels.

제안 방법

• Propose DiffUCD with two main modules: Semantic Correlation Diffusion Model (SCDM) and Cross-Temporal Contrastive Learning (CTCL), plus a Change Detection head.
• Train SCDM in a pretraining stage using large unlabeled HSI-CD samples to capture spectral-spatial semantic correlations.
• Freeze SCDM and train CTCL with pseudo-labels to align unchanged spectral features across time via contrastive learning.
• Use a fusion module to combine SCDM-derived features and CTCL-derived invariant features for pixel-wise change detection.
• Optimize with a combined loss including diffusion-based reconstruction, CTCL contrastive loss, and change detection loss.
• Demonstrate denoising and reconstruction of original semantic features through the diffusion process and analyze timestamp t effects.

실험 결과

연구 질문

• RQ1 Can a diffusion-based model be adapted effectively for unsupervised HSI-CD by exploiting semantic correlations?
• RQ2 How can cross-temporal spectral differences caused by environmental changes be learned as invariant features?
• RQ3 Does integrating SCDM with CTCL and a fusion change head outperform existing unsupervised methods on public datasets?

주요 결과

Method	OA	KC	F1
Santa Barbara (Ours)	96.87	93.41	95.97
Bay Area (Ours)	96.35	92.67	96.57
Hermiston (Ours)	95.47	86.69	89.58

• DiffUCD achieves state-of-the-art OA, KC, and F1 among unsupervised HSI-CD methods on Santa Barbara, Bay Area, and Hermiston datasets (OA up to 96.87%, KC up to 93.41%, F1 up to 95.97% on Santa Barbara).
• DiffUCD outperforms prior unsupervised methods by approximately 5.73% OA, 11.93% KC, and 7.17% F1 on Santa Barbara.
• With comparable labeled data, DiffUCD shows competitive performance to supervised methods, sometimes surpassing them in OA/KC/F1.
• Ablation studies show significant gains from incorporating SCDM and CTCL, with DiffUCD yielding substantial improvements over the base model on all three datasets.
• t-SNE visualizations indicate DiffUCD features are more separable, reflecting improved intra-/inter-class discrimination.
• Qualitative results illustrate effective noise removal and reconstruction of original semantic spectral-spatial features during diffusion.