SCDM-HTD: Spectral-Wise Conditional Diffusion Model for Hyperspectral Target Detection

Dunbin Shen; Wenfeng Kong; Jianjun Liu; Xiaorui Ma; Hongyu Wang

cs.CVOct 23, 2026

SCDM-HTD: Spectral-Wise Conditional Diffusion Model for Hyperspectral Target Detection

Authors: Dunbin Shen, Wenfeng Kong, Jianjun Liu, Xiaorui Ma, Hongyu Wang

Organizations: School of Information and Communication Engineering, Dalian University of Technology, Dalian 116024, China · School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China

Abstract

Hyperspectral target detection (HTD) aims to locate and recognize objects at a fine-grained scale, playing a pivotal role in remote sensing. Despite significant progress, current HTD methods primarily focus on feature comparison while neglecting distribution learning for background-target separation, which can limit detection robustness in practical applications. This paper addresses this issue by introducing a spectral-wise conditional diffusion model that reformulates HTD as a task of generating a detection map. Specifically, we treat background and target confidence scores as training samples and use the corresponding pixel spectra as guiding conditions to capture the distributional discrepancy between background and target through diffusion and denoising. To enhance practical applicability, we propose a lightweight conditional denoising network based on multilayer perceptron to map noisy confidence scores to their ideal counterparts and adopt a deterministic reverse non-Markov chain to skip sampling steps, thereby accelerating the generation process. Furthermore, we introduce a composite loss to improve background-target separation. This loss integrates constrained energy minimization loss to enhance target saliency and suppress the background by maximizing the target response while minimizing average output energy, alongside binary cross-entropy loss to classify background and target using pixel-wise labels. Experimental results on four benchmark datasets demonstrate the effectiveness of the proposed method, with improvements in both quantitative metrics and qualitative visual effects.

Figures & tables

Fig. 1: Overall framework of the proposed SCDM-HTD.

Fig. 2: Forward and reverse processes of SCDM-HTD: The forward diffusion progressively adds Gaussian noise to the binary confidence maps, gradually corrupting them. In contrast, the reverse sampling process iteratively denoises the maps to recover the confidence scores, conditioned on the corresponding pixel spectra.

Layer		Group size	Input size	Output size
Initial layer	Linear	-	$B\times 1$	$B\times 128$
Time embedding	Embed	-	$B\times 1$	$B\times 128$
Conditional embedding	Linear	-	$B\times L$	$B\times 128$
MLP	Linear+SiLU	-	$B\times 128$	$B\times 512$
MLP	Linear	-	$B\times 512$	$B\times 512$
Residual block	GroupNorm+SiLU	32	$B\times 128$	$B\times 128$

TABLE I: Detailed configuration of CDNet

Panel of Figure 5 — Fig. 4: Visualized results of the competing methods on (a) MULL Gulfport, (b) HYDICE, (c) Airport-4, and (d) Urban-2.

Panel of Figure 6 — Fig. 5: ROC curves of the competing methods on (a) MULL Gulfport, (b) HYDICE, (c) Airport-4, and (d) Urban-2. From left to right: 3D ROC curve, 2D ROC curve of $\rm(P_{f},P_{d})$ , 2D ROC curve of $\rm(\tau,P_{f})$ , and 2D ROC curve of $\rm(\tau,P_{d})$ .

Panel of Figure 7 — Fig. 6: Background-target separability diagrams of competing methods on four datasets. (a) MULL Gulfport. (b) HYDICE. (c) Airport-4. (d) Urban-2.

Dataset		OSP	CEM	CSCR	WDCCR	BLTSC	HTD-Net	TSTTD	OS-VAE	SCDM-HTD
MULL Gulfport	$\rm{AUC}_{(P_{f},P_{d})}$	0.9690	0.8257	0.9993	0.9184	0.6943	0.9719	0.9958	0.7501	0.9999
	$\rm{AUC}_{(\tau,P_{d})}$	0.8553	0.6493	0.8646	0.8981	0.2435	0.9314	0.2666	0.3934	0.6828
	$\rm{AUC}_{(\tau,P_{f})}$	0.2724	0.4162	0.2960	0.7510	0.0028	0.3960	0.0122	0.0873	0.0039
	$\rm{AUC}_{\rm{OA}}$	1.5519	1.0588	1.5679	1.0655	0.9350	1.5073	1.2502	1.0562	1.6788
	$\rm{AUC}_{\rm{SNPR}}$	3.1399	1.5601	2.9209	1.1959	86.9643	2.3520	21.8525	4.5063	175.0769
	Running time	0.0090	0.0061	0.1357	2.2048	1.7499	4.0861	1.7609	0.4708	0.3357

TABLE II: AUC SCORES AND RUNNING TIME (IN SECONDS) OF COMPETING METHODS ON THE FOUR DATASETS. BOLD REPRESENTS THE BEST RESULT WHILE UNDERLINED THE SECOND.

Panel of Figure 9 — Fig. 7: Effect of parameters. (a) $\mu$ . (b) $\lambda$ .

Loss	MULL Gulfport	HYDICE	Airport-4	Urban-2
Diffusion loss	0.7804	0.9947	0.9669	0.9499
Separation loss	0.9999	0.9998	0.9986	0.9969

TABLE III: Effectiveness of background-target separation loss on four datasets. Bold represents the best result.

Datasets	CEM loss	BCE loss	${\rm AUC_{(P_{f},P_{d})}}$
MULL Gulfport		✓	0.9996
	✓		0.9993
	✓	✓	0.9999
HYDICE		✓	0.8646
	✓		0.9996
	✓	✓	0.9998

TABLE IV: Effectiveness of CEM loss and BCE loss on four datasets. Bold represents the best result.

Dataset	DDPM			Reverse sampling acceleration
	$\rm{AUC}_{(P_{f},P_{d})}$	Steps	Time (in seconds)	$\rm{AUC}_{(P_{f},P_{d})}$	Steps	Time (in seconds)
MULL Gulfport	1.0000	500	1.7158	0.9999	5	0.3357
HYDICE	0.9998	500	4.1429	0.9998	5	0.3552
Airport-4	0.9987	500	4.7545	0.9986	5	0.3845
Urban-2	0.9969	500	4.8850	0.9969	5	0.3826

TABLE V: Effectiveness of reverse sampling acceleration on four datasets. Bold represents the best result.

Hyperspectral Imaging

Hyperspectral imaging captures spatial images across many narrow wavelength bands to identify materials and support detection, mapping, and analysis.

See all 133 papers

Extending Reconstruction: Reconstruction-to-Vector Diffusion for Hyperspectral Anomaly Detection
Apr 13, 2026Jijun Xiang, Tao Wang, Jiayi Wang +3Anomaly Detection Hyperspectral Imaging
the Rocket Force University of Engineering, Xi’an 710025, China
Physics-Aligned Spectral Mamba: Decoupling Semantics and Dynamics for Few-Shot Hyperspectral Target Detection
Apr 7, 2026Luqi Gong, Qixin Xie, Yue Chen +4Hyperspectral Imaging Parameter-Efficient Fine-Tuning
Research Center for Space Computing System, Zhejiang Lab, Hangzhou 311121, China · State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China
High-Dimensional Noise to Low-Dimensional Manifolds: A Manifold-Space Diffusion Framework for Degraded Hyperspectral Image Classification
Apr 29, 2026Boxiang Yang, Ning Chen, Xia Yue +6Manifold Learning Hyperspectral Image Classification
Institute of Remote Sensing and Geographic Information System, Peking University, Beijing 100871, China · Beijing Key Laboratory of Spatio-temporal Perception and Urban Resilience, Peking University, Beijing 100871, China · School of Computer Science and Engineering, Central South University, Changsha 410083, China +4

Apr 13, 2026J/K move · Enter open · S save

Extending Reconstruction: Reconstruction-to-Vector Diffusion for Hyperspectral Anomaly Detection

Jijun Xiang, Tao Wang, Jiayi Wang, Pengxiang Wang+2

the Rocket Force University of Engineering, Xi’an 710025, China

Anomaly Detection Hyperspectral Imaging Reconstruction-Based Anomaly Detection

Open paper AskPDF arXiv

While Hyperspectral Anomaly Detection (HAD) excels at identifying sparse targets in complex scenes, existing models remain trapped in a scalar "reconstruction-as-endpoint" paradigm. This reliance on ambiguous scalar residuals consistently triggers sub-pixel anomaly vanishing during spatial downsampling, alongside severe confirmation bias when unpurified anomalies corrupt training weights. In this paper, we propose Reconstruction-to-Vector Diffusion (R2VD), which fundamentally redefines reconstruction as a manifold purification origin to establish a novel residual-guided generative dynamics paradigm. Our framework introduces a four-stage pipeline: (1) a Physical Prior Extraction (PPE) stage that mitigates early confirmation bias via dual-stream statistical guidance; (2) a Guided Manifold Purification (GMP) stage utilizing an OmniContext Autoencoder (OCA) to extract purified residual maps while preserving fragile sub-pixel topologies; (3) a Residual Score Modeling (RSM) stage where a Diffusion Transformer (DiT), guarded by a Physical Spectral Firewall (PSF), effectively isolates cross-spectral leakage; and (4) a Vector Dynamics Inference (VDI) stage that robustly decouples targets from backgrounds by evaluating high-dimensional vector interference patterns instead of conventional scalar errors. Comprehensive evaluations on eight datasets confirm that R2VD establishes a new state-of-the-art, delivering exceptional target detectability and background suppression. The code is available at https://github.com/Bondojijun/R2VD.

Abstract

Hyperspectral Imaging

Explore similar work

Extending Reconstruction: Reconstruction-to-Vector Diffusion for Hyperspectral Anomaly Detection

Physics-Aligned Spectral Mamba: Decoupling Semantics and Dynamics for Few-Shot Hyperspectral Target Detection

High-Dimensional Noise to Low-Dimensional Manifolds: A Manifold-Space Diffusion Framework for Degraded Hyperspectral Image Classification