SpecEStop: Self-Supervised Hyperspectral Mixed Noise Removal via Deep Spectral Prior.

Area of Science:

• Remote Sensing
• Image Processing
• Artificial Intelligence

Background:

• Hyperspectral remote sensing images (HSIs) are corrupted by mixed noise (Gaussian, stripe, impulse) due to various environmental and sensor factors.
• Existing denoising methods struggle with band-dependent, diverse noise distributions and often require clean/noisy data pairs or manual tuning.
• These limitations hinder the effective analysis and application of real-world HSIs.

Purpose of the Study:

• To develop a fully self-supervised framework for hyperspectral image denoising that requires only a single noisy HSI for training.
• To address the challenge of mixed noise removal in HSIs without relying on clean training data.
• To improve the quality and usability of hyperspectral remote sensing data for various applications.

Main Methods:

• Proposing SpecEStop, a self-supervised spectral vector denoising framework utilizing a deep spectral prior.
• Exploiting neural network spectral bias for adaptive early stopping to prevent learning non-Gaussian noise.
• Designing a network architecture to preserve Gaussian noise properties in the latent space for subsequent removal.

Main Results:

• SpecEStop effectively suppresses complex, mixed noise patterns in hyperspectral images through stage-wise denoising.
• The framework achieves robust performance across diverse real-world hyperspectral remote sensing scenarios.
• Validated effectiveness of the self-supervised approach without requiring clean/noisy image pairs for training.

Conclusions:

• SpecEStop offers a powerful, adaptable solution for hyperspectral image denoising, overcoming limitations of existing methods.
• The self-supervised, spectral prior-based approach enables high-quality noise removal from single noisy HSIs.
• This framework has significant implications for advancing hyperspectral data analysis and applications in remote sensing.