Tensor Multi-Subspace Representation for Remote Sensing Image Mixed Noise Removal.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Remote sensing image (RSI) denoising is crucial for data quality.
• Current methods often assume a single subspace, which is insufficient for complex RSIs.
• Wavelength differences and temporal variations necessitate advanced denoising approaches.

Purpose of the Study:

• To propose a novel Tensor Multi-Subspace Representation (TenMSR) for mixed noise removal in RSIs.
• To accurately characterize the intrinsic multi-subspace structure of RSI data.
• To enhance denoising performance by addressing wavelength and temporal variability.

Main Methods:

• Developed TenMSR to represent RSI data within multiple tensor subspaces.
• Introduced a nonlinear transform-based 3-D tensor nuclear norm for low-rank characterization.
• Implemented an algorithm using the proximal alternating minimization (PAM) framework for model optimization.

Main Results:

• TenMSR precisely describes wavelength differences and temporal variability in RSIs.
• The method achieves a more compact image distribution within the tensor multi-subspace.
• Experimental results demonstrate superior performance compared to state-of-the-art single subspace methods.

Conclusions:

• TenMSR effectively removes mixed noise from RSIs by leveraging multi-subspace characteristics.
• The proposed method offers a significant advancement over traditional single subspace denoising techniques.
• TenMSR provides a robust framework for processing complex remote sensing data.