Abstract
Hyperspectral remote sensing images frequently suffer from varying degrees of degradation, which can severely hinder downstream applications like classification and target detection. Accurate noise level estimation is essential for effective data preprocessing. Because noise levels often fluctuate significantly across different spectral bands, this paper introduces a noise estimation algorithm based on low-rank representation (LRR). By exploiting the inherent spectral correlation between bands, the proposed method constructs a low-rank model to estimate band-specific noise levels and identify "useless" or invalid bands. Experimental results on both simulated and real-world hyperspectral datasets demonstrate that the algorithm provides accurate and robust noise estimation, achieving a 0% miss rate for invalid bands and over 80% detection accuracy within appropriate threshold ranges.
Introduction
Hyperspectral sensors capture data across a vast range of the electromagnetic spectrum, providing rich information for remote sensing. However, this richness comes with a significant challenge: non-uniform noise. Factors such as atmospheric interference (e.g., water vapor absorption) and sensor limitations mean that while one band might be crystal clear, the next could be completely obscured by Gaussian noise, salt-and-pepper artifacts, or "striping."
Existing noise estimation techniques often fall into two traps: they either assume noise is uniform across all bands—which is rarely the case in practice—or they process each band independently, ignoring the massive amount of redundant information shared between adjacent spectral channels. This research aims to bridge this gap by creating a model that is globally aware of spectral correlations but locally sensitive to band-specific noise.
Methodology
The core of the proposed approach is to treat the hyperspectral data cube as a matrix and decompose it into two components: a low-rank "clean" signal and a noise matrix.
Low-Rank Modeling
Because adjacent spectral bands are highly correlated, the "true" underlying signal can be represented by a low-rank matrix. The 3D HSI cube is first flattened into a 2D matrix where each column represents a spectral band.
The $L_{2,1}$ Norm Innovation
Standard low-rank methods often use the $L_1$ norm for sparse noise or the $L_2$ norm for Gaussian noise. However, these do not account for the fact that noise levels differ between bands. The authors employ the $L_{2,1}$ norm (group sparsity norm) to regularize the noise term. This mathematical choice allows the model to recognize that noise is "grouped" by band (column-wise), making it significantly more effective at capturing band-specific noise levels. This dual approach—utilizing inter-band correlation for global processing while preserving local band differences—is key to the method's robustness.
Optimization and Assessment
The optimization problem is solved using the Inexact Augmented Lagrange Multiplier (IALM) method, which efficiently handles the complex constrained optimization. Once the noise matrix is estimated, the $L_2$ norm of each column provides a direct "noise score" for that band. If this score exceeds a predefined threshold, the band is flagged as "useless" and can be removed from further processing.
Experimental Results
The researchers evaluated their approach using both simulated and real-world datasets:
- Simulated Data: Using a subset of the HYDICE Washington DC National Mall dataset, various noise types (Gaussian, salt-and-pepper, and stripe noise) were added to clean bands. The proposed method's noise estimates showed a much tighter correlation with actual noise levels compared to standard methods like VBM3D or wavelet-based estimation.
- Real-World Data:
- Salinas (AVIRIS): Used to verify the correlation between estimated noise and visual image quality. Bands labeled as "low noise" were visually sharp, while "high noise" bands were visibly degraded.
- Indian Pines (AVIRIS): Specifically used to test the detection of "useless" bands, such as those affected by water absorption. The algorithm successfully identified these bands with a 0% miss rate and an overall detection accuracy higher than 80%.
Discussion and Key Insights
The primary innovation of this work lies in using the low-rank assumption to capture global spectral commonalities, which effectively "pushes" band-specific differences and noise into the residual matrix. By using the $L_{2,1}$ norm, the method gains robustness against "bursty" noise like salt-and-pepper or stripes, which often defeat traditional variance-based methods. This highlights that noise estimation is not just a precursor to denoising, but a critical diagnostic tool for determining data reliability.
Conclusion and Future Work
The proposed LRR-based method offers a robust solution for non-uniform noise estimation in HSI. However, it requires precise geometric alignment of the hyperspectral images; if pixels across bands do not represent the same physical location, the low-rank assumption fails. Future research could focus on integrating this noise estimation directly into automated HSI processing pipelines to enhance the performance of subsequent classification and target detection algorithms.
Summary updated