Spectral Reflectance Estimation Using Gaussian Processes and Combination Kernels.

Area of Science:

• Optics and Photonics
• Machine Learning
• Remote Sensing

Background:

• Hyperspectral reflectance factor estimation is crucial for various applications.
• Traditional methods often require detailed sensor and light spectral characteristics.
• Gaussian process regression offers a probabilistic framework for regression tasks.

Purpose of the Study:

• To develop and evaluate novel Gaussian process regression models for hyperspectral reflectance factor estimation.
• To assess the performance of these models using multispectral and trichromatic measurements.
• To investigate the effectiveness of learning-based approaches without prior spectral information.

Main Methods:

• Construction of new Gaussian process (GP) models for spectral estimation.
• Utilizing marginal likelihood optimization for parameter selection and model evaluation.
• Employing anisotropic radial and combination kernels for process covariance.
• Applying input data transformations for pre-processing.
• Exploring spectral subspace coordinate learning.

Main Results:

• Several new GP models demonstrated improved spectral accuracy compared to previous kernel-based models.
• The proposed models achieved better performance on both simulated and real-world data.
• A versatile model combining spectral subspace coordinate learning and combination kernels showed efficient optimization via marginal likelihood.
• Preliminary results indicated the models provide uncertainty estimates.

Conclusions:

• Novel Gaussian process models offer enhanced accuracy for hyperspectral reflectance estimation.
• The learning-based approach obviates the need for sensor and light spectral data.
• The developed models, particularly the subspace coordinate learning variant, are promising for spectral estimation and uncertainty quantification.