Trainable spectral difference learning with spatial starting for hyperspectral image denoising.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral images (HSI) suffer from low signal-to-noise ratios due to limited energy and illumination.
• Noise significantly degrades HSI analysis performance.
• Existing denoising methods often distort critical spectral information.

Purpose of the Study:

• To develop an HSI denoising method that preserves spectral information.
• To improve the accuracy and efficiency of HSI analysis.

Main Methods:

• A novel hyperspectral image denoising method (HDnTSDL) is proposed.
• It utilizes trainable spectral difference learning with spatial initialization.
• A key band is denoised first, guiding the reconstruction of other bands via a CNN with learned non-linear functions.

Main Results:

• HDnTSDL demonstrates superior performance in spatial recovery and spectral preservation compared to state-of-the-art methods.
• The method achieves significant noise reduction across diverse indoor and outdoor HSI datasets.
• Experimental results validate the effectiveness on five databases.

Conclusions:

• The proposed HDnTSDL method effectively denoises hyperspectral images while preserving spectral fidelity.
• It offers a computationally efficient alternative to existing HSI denoising techniques.
• HDnTSDL enhances the reliability of subsequent HSI analysis and interpretation.