Emulation as an Accurate Alternative to Interpolation in Sampling Radiative Transfer Codes.

Area of Science:

• Earth and Atmospheric Sciences
• Computational Modeling
• Remote Sensing

Background:

• Radiative transfer models (RTMs) are crucial for simulating light interactions with Earth's surface and atmosphere.
• Current RTM practices often rely on interpolating sparse look-up tables (LUTs) due to high computational costs.
• The accuracy of common interpolation methods for RTM data is questionable.

Purpose of the Study:

• To evaluate the accuracy of emulation methods versus traditional interpolation for RTM spectral data reconstruction.
• To compare the performance of various interpolation and emulation techniques.

Main Methods:

• Two experiments were conducted using PROSAIL (canopy level) and MODTRAN (top-of-atmosphere level).
• Interpolation methods tested: nearest-neighbor, inverse distance weighting, piecewise linear.
• Emulation methods tested: Gaussian Process Regression (GPR), kernel ridge regression, neural networks.
• Performance was assessed against a dense reference LUT.

Main Results:

• Emulation methods significantly outperformed classical interpolation methods in spectral output accuracy.
• Gaussian Process Regression (GPR) emulation achieved up to ten times greater accuracy than the best interpolation method.
• Emulation methods demonstrated competitive speed, comparable to faster interpolation techniques.

Conclusions:

• Emulation, particularly GPR, provides a more accurate and efficient alternative to interpolation for RTM spectral data.
• Statistical learning-based emulation can accelerate the analysis of complex radiative transfer simulations.