SpecResNet: Hyperspectral Image Compression via Hybrid Residual Learning and Spectral Calibration.
Saved in:
| Title: | SpecResNet: Hyperspectral Image Compression via Hybrid Residual Learning and Spectral Calibration. |
|---|---|
| Authors: | Saeed, Fahad1,2 (AUTHOR), Liu, Shumin1,2,3 (AUTHOR) liu_shumin@nwpu.edu.cn, Chen, Jie1,3 (AUTHOR) |
| Source: | Remote Sensing. Apr2026, Vol. 18 Issue 7, p1074. 17p. |
| Subjects: | Image compression, Remote sensing, Deep learning, Autoencoders, Wavelength measurement, Optimization algorithms, Rate distortion theory |
| Abstract: | Highlights: What are the main findings? The proposed SpecResNet achieves superior rate–distortion performance across five benchmark hyperspectral datasets, with consistent improvements in PSNR, MS-SSIM, and SAM compared to existing methods. The hybrid residual block reduces computational complexity by more than 10× in FLOPs and over 3× in parameters compared to standard residual designs, while the spectral calibration block provides measurable gains in spectral accuracy, improving overall reconstruction quality. What is the implication of the main finding? Efficient architectural redesign significantly reduces computational cost in 3D hyperspectral compression while preserving high reconstruction performance, demonstrating that complexity can be lowered without sacrificing fidelity. Explicit spectral refinement improves spectral accuracy and enables a better complexity–distortion trade-off, making the framework more suitable for practical remote sensing hyperspectral imaging applications. Hyperspectral imaging provides rich spatial–spectral information but generates huge data volumes, posing significant challenges for storage, transmission, and real-time processing in remote sensing applications. In this study, we propose SpecResNet, a 3D autoencoder-based model for hyperspectral image compression. This framework introduces hybrid residual blocks for preserving representational power and a spectral calibration (SC) block to enhance spectral fidelity. It also uses Squeeze-and-Excitation (SE) blocks for adaptive feature recalibration. Our model obtains different compression operating points by varying model capacity, with bitrate emerging implicitly from the learned latent representations. Experiments on several benchmark datasets show that SpecResNet surpasses the performance of existing frameworks on most datasets in terms of PSNR, MS-SSIM, and SAM, demonstrating its strong potential. Our results suggest that SpecResNet offers a promising trade-off for efficient hyperspectral image compression, with potential for further refinement in complex scenes. [ABSTRACT FROM AUTHOR] |
| Copyright of Remote Sensing is the property of MDPI and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Highlights: What are the main findings? The proposed SpecResNet achieves superior rate–distortion performance across five benchmark hyperspectral datasets, with consistent improvements in PSNR, MS-SSIM, and SAM compared to existing methods. The hybrid residual block reduces computational complexity by more than 10× in FLOPs and over 3× in parameters compared to standard residual designs, while the spectral calibration block provides measurable gains in spectral accuracy, improving overall reconstruction quality. What is the implication of the main finding? Efficient architectural redesign significantly reduces computational cost in 3D hyperspectral compression while preserving high reconstruction performance, demonstrating that complexity can be lowered without sacrificing fidelity. Explicit spectral refinement improves spectral accuracy and enables a better complexity–distortion trade-off, making the framework more suitable for practical remote sensing hyperspectral imaging applications. Hyperspectral imaging provides rich spatial–spectral information but generates huge data volumes, posing significant challenges for storage, transmission, and real-time processing in remote sensing applications. In this study, we propose SpecResNet, a 3D autoencoder-based model for hyperspectral image compression. This framework introduces hybrid residual blocks for preserving representational power and a spectral calibration (SC) block to enhance spectral fidelity. It also uses Squeeze-and-Excitation (SE) blocks for adaptive feature recalibration. Our model obtains different compression operating points by varying model capacity, with bitrate emerging implicitly from the learned latent representations. Experiments on several benchmark datasets show that SpecResNet surpasses the performance of existing frameworks on most datasets in terms of PSNR, MS-SSIM, and SAM, demonstrating its strong potential. Our results suggest that SpecResNet offers a promising trade-off for efficient hyperspectral image compression, with potential for further refinement in complex scenes. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 20724292 |
| DOI: | 10.3390/rs18071074 |
