Unsupervised Adaptation Learning for Real Multiplatform Hyperspectral Image Denoising.

Area of Science:

• Remote Sensing
• Image Processing
• Artificial Intelligence

Background:

• Real hyperspectral images (HSIs) suffer from significant noise, limiting their practical applications.
• Existing transfer learning methods for hyperspectral denoising often depend on image priors and fail to preserve background details.
• Current hyperspectral denoising techniques struggle with generalizability across diverse datasets and noise types.

Purpose of the Study:

• To propose an unsupervised adaptation learning (UAL)-based hyperspectral denoising network (UALHDN) for improved HSI noise reduction.
• To develop a framework that learns general image priors and adapts them to real HSIs while preserving background consistency.
• To address the limitations of existing methods by eliminating the need for hand-crafted priors and improving model generalizability.

Main Methods:

• Introduced a UALHDN framework incorporating a spatial-spectral residual denoiser, a global modeling discriminator, and a hyperspectral discrete representation learning scheme.
• Pretrained the denoiser and discriminator on synthetic noisy-clean HSI pairs.
• Fine-tuned the denoiser on real multiplatform HSIs using unsupervised spatial-spectral and background consistency losses, alongside discrete representation learning for semantic feature extraction.

Main Results:

• The UALHDN framework demonstrated superior denoising performance compared to state-of-the-art methods on real-world HSIs.
• Experiments validated the framework's applicability and generalizability across diverse datasets from various platforms (UAV, airborne, spaceborne) and sensors, including Martian data.
• The unsupervised adaptation approach effectively learned deep priors specific to real HSIs, enhancing noise removal while maintaining image fidelity.

Conclusions:

• The proposed UALHDN offers a robust and effective solution for hyperspectral image denoising, overcoming limitations of previous approaches.
• Unsupervised adaptation learning is a promising strategy for hyperspectral image processing, enabling high-fidelity denoising without manual priors.
• The UALHDN framework shows significant potential for various applications requiring clean hyperspectral data across different acquisition platforms.