uTDSP

Transformer-based Diffusion and Spectral Priors Model For Hyperspectral Pansharpening

📄 [Paper Link (IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 2025)]

Hyperspectral Pansharpening with Transformer-based Spectral Diffusion Priors

📄 [Paper Link (The IEEE/CVF Winter Conference on Applications of Computer Vision 2025)]

Authors:
Hongcheng Jiang
Zhiqiang Chen

---

---

🔍 Overview

The goal is to reconstruct a high-resolution hyperspectral image (HR-HSI) by fusing a low-resolution hyperspectral image (LR-HSI) and a high-resolution panchromatic image (HR-PCI). Unlike conventional methods that rely on paired HR-HSI ground truth, uTDSP is entirely unsupervised, leveraging spectral priors and a transformer-based diffusion model to guide the reconstruction process.

---

🧠 Key Features

• 🎯 Unsupervised Learning: Learns directly from LR-HSI and HR-PCI without requiring any ground-truth HR-HSI.
• 🌀 Spectral Diffusion Process: Incorporates a transformer-based denoiser within a diffusion framework.
• 🧩 Spectral Prior Integration: Enforces spectral consistency using priors extracted from the LR-HSI.
• ⚖️ Adaptive Loss Balancing: Combines spectral fidelity loss and diffusion consistency for robust reconstruction.
• 🏆 SOTA Results: Achieves superior performance across multiple benchmark datasets.

---

📈 Performance Gains

uTDSP achieves consistent PSNR improvements across diverse airborne and satellite datasets, outperforming both supervised and unsupervised baselines.

✅ Airborne Datasets

• Chikusei

  • uTDSP: 26.86 dB
  • Best prior method (DDLPS*): 26.85 dB
  • 🔺 +0.01 dB (+0.04%)

• Indian Pines

  • uTDSP: 25.79 dB
  • Best prior method (DIP-HyperKite): 25.15 dB
  • 🔺 +0.64 dB (+2.54%)

• PaviaC

  • uTDSP: 28.53 dB
  • Best prior method (DDLPS*): 27.52 dB
  • 🔺 +1.01 dB (+3.67%)

• PaviaU

  • uTDSP: 30.68 dB
  • Best prior method (GPPNN): 29.86 dB
  • 🔺 +0.82 dB (+2.75%)

✅ Satellite Datasets

• Botswana

  • uTDSP: 31.61 dB
  • Best prior method (DIP-HyperKite): 30.24 dB
  • 🔺 +1.37 dB (+4.53%)

• ZY1-02D

  • uTDSP: 31.22 dB
  • Best prior method (SFIM*): 28.23 dB
  • 🔺 +2.99 dB (+10.59%)

---

📌 Summary

uTDSP consistently improves PSNR by:

• +0.01–1.01 dB (up to +3.7%) on airborne datasets
• +1.37–2.99 dB (up to +10.6%) on satellite datasets

These results confirm uTDSP’s strong generalization and superior reconstruction quality across sensing platforms—all achieved without supervision.

🔄 Diffusion Process Illustration

---

🧠 Network Architecture

📊 Band-wise PSNR Comparison

The following plots illustrate the PSNR values for each spectral band across six benchmark datasets, highlighting the spectral fidelity of uTDSP compared to existing methods.

🛰️ Botswana

🛰️ Chikusei

🛰️ Pavia Center

🛰️ Pavia University

🛰️ Indian Pines

🛰️ Ziyuan-1

---

🖼️ Visual Results

📷 Airborne datasets

📷 Satellite Datasets

---

📊 Quantitative Results on Airborne Datasets

Metrics: PSNR ↑ (higher is better), SAM ↓, ERGAS ↓
( denotes an unsupervised method)*

---

🛰️ Chikusei

Method	PSNR	SAM	ERGAS
DBDENet	25.02	6.5243	4.2316
DHP-DARN	25.24	6.0044	3.9208
DIP-HyperKite	25.63	5.4180	3.7059
DMLD-Net	25.28	6.9856	4.1170
GPPNN	25.17	6.5704	4.1423
HyperPNN	25.34	5.7174	3.8096
DDLPS*	26.85	5.3557	3.7616
GSA*	24.21	6.2670	5.3903
Indusion*	22.29	5.4171	5.0367
PLRDiff*	26.18	6.3831	3.5699
SFIM*	25.54	5.4171	4.0868
uTDSP*	26.86	5.9152	3.3178

---

🛰️ Indian Pines

Method	PSNR	SAM	ERGAS
DBDENet	23.66	3.4962	1.6389
DHP-DARN	23.97	3.6511	2.1060
DIP-HyperKite	25.15	3.6619	1.4592
DMLD-Net	24.03	3.9927	1.7721
GPPNN	24.80	4.4834	1.5553
HyperPNN	24.60	3.8811	1.5532
DDLPS*	17.72	4.5282	10.0461
GSA*	20.44	4.0885	5.0318
Indusion*	4.83	3.8504	14.8366
PLRDiff*	8.10	10.6969	19.8362
SFIM*	24.42	3.8504	1.5483
uTDSP*	25.79	3.5621	1.2763

---

🛰️ PaviaC

Method	PSNR	SAM	ERGAS
DBDENet	23.63	18.5981	6.7944
DHP-DARN	26.70	12.4018	4.7917
DIP-HyperKite	26.48	12.5846	4.9198
DMLD-Net	26.03	16.8061	5.1832
GPPNN	27.37	11.2643	4.4916
HyperPNN	26.10	17.5919	5.1762
DDLPS*	27.52	10.1478	4.3781
GSA*	25.30	10.4678	5.6518
Indusion*	25.84	10.4645	5.8683
PLRDiff*	27.39	11.6904	4.6245
SFIM*	24.99	10.4488	5.9011
uTDSP*	28.53	10.4176	4.2664

---

🛰️ PaviaU

Method	PSNR	SAM	ERGAS
DBDENet	28.84	6.5032	2.7593
DHP-DARN	29.27	6.8826	2.5350
DIP-HyperKite	29.30	6.0972	2.5114
DMLD-Net	28.66	6.8624	2.7985
GPPNN	29.86	6.0788	2.4812
HyperPNN	28.96	6.7555	2.6472
DDLPS*	27.81	7.1405	4.3781
GSA*	26.47	7.2522	2.9323
Indusion*	25.82	7.8229	4.1539
PLRDiff*	28.57	7.5217	2.9453
SFIM*	25.66	7.8229	3.8125
uTDSP*	30.68	6.8019	2.4660

🛰️ Quantitative Results on Satellite Datasets

Metrics: PSNR ↑ (higher is better), SAM ↓, ERGAS ↓
( denotes an unsupervised method)*

---

🛰️ Botswana

Method	PSNR	SAM	ERGAS
DBDENet	22.84	8.5207	11.3979
DHP-DARN	28.85	4.9084	2.8164
DIP-HyperKite	30.24	4.8305	2.1305
DMLD-Net	26.87	6.5379	3.7552
GPPNN	26.44	8.6439	3.8965
HyperPNN	29.83	4.9803	2.2254
DDLPS*	22.27	6.9539	17.5198
GSA*	23.80	6.2035	11.6626
Indusion*	15.30	5.4225	9.7633
PLRDiff*	17.84	15.1475	9.0164
SFIM*	26.81	5.4225	2.7995
uTDSP*	31.61	3.7777	1.9155

---

🛰️ ZY1-02D

Method	PSNR	SAM	ERGAS
DBDENet	11.39	22.3445	29.8753
DHP-DARN	14.71	7.9514	24.0358
DIP-HyperKite	19.73	2.1563	3.8904
DMLD-Net	13.41	13.4045	11.3059
GPPNN	14.35	12.1717	10.6018
HyperPNN	19.42	2.7872	5.5515
DDLPS*	26.03	1.8212	5.0919
GSA*	16.60	4.5868	20.0515
Indusion*	14.53	1.7358	6.7524
PLRDiff*	26.77	2.5636	2.7032
SFIM*	28.23	1.7358	1.6604
uTDSP*	31.22	1.7466	1.6917

📚 Citation

If you find this work helpful in your research, please cite:

@ARTICLE{11085108,
  author={Jiang, Hongcheng and Chen, ZhiQiang},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, 
  title={Transformer-based Diffusion and Spectral Priors Model For Hyperspectral Pansharpening}, 
  year={2025},
  volume={},
  number={},
  pages={1-17},
  keywords={Hyperspectral imaging;Pansharpening;Diffusion models;Transformers;Estimation;Bayes methods;Noise reduction;Image reconstruction;Earth;Degradation;Hyperspectral imaging;pansharpening;spectral priors;diffusion model;transformer;remote sensing},
  doi={10.1109/JSTARS.2025.3590685}}

@inproceedings{jiang2025hyperspectral,
  title={Hyperspectral Pansharpening with Transformer-Based Spectral Diffusion Priors},
  author={Jiang, Hongcheng and Chen, ZhiQiang},
  booktitle={Proceedings of the Winter Conference on Applications of Computer Vision},
  pages={581--590},
  year={2025}
}

📬 Contact

If you have any questions, feedback, or collaboration ideas, feel free to reach out:

• 💻 Website: jianghongcheng.github.io
• 📧 Email: [email protected]
• 🏫 Affiliation: University of Missouri–Kansas City (UMKC)