Systematic Assessment of MODTRAN Emulators for Atmospheric Correction.

Area of Science:

• Earth and Atmospheric Sciences
• Remote Sensing
• Computational Modeling

Background:

• Atmospheric radiative transfer models (RTMs) are crucial for simulating light propagation but are computationally intensive.
• Traditional lookup table (LUT) interpolation methods require large datasets and significant computation time.
• Emulation offers a computationally efficient alternative by approximating RTM outputs with statistical models.

Purpose of the Study:

• To systematically assess factors influencing the precision of emulating MODTRAN for atmospheric correction.
• To determine the optimal regression algorithm, training data size, dimensionality reduction technique, and spectral resolution for accurate emulation.
• To validate the accuracy and efficiency of emulators for satellite data processing.

Main Methods:

• Evaluated multiple regression algorithms, including Gaussian processes regression (GPR).
• Investigated the impact of training database size and dimensionality reduction (DR) methods like principal component analysis (PCA).
• Assessed performance across different spectral resolutions and number of principal components.

Main Results:

• Gaussian processes regression (GPR) demonstrated the highest accuracy among tested emulators.
• Principal component analysis (PCA) with approximately 20 components proved to be a robust DR method.
• GPR emulators, trained on 1000 samples, achieved relative errors below 1% (95th percentile) for spectral data reconstruction.
• Processing time was reduced from days to minutes, maintaining sufficient accuracy.

Conclusions:

• Emulators, particularly GPR, provide a computationally efficient and accurate solution for atmospheric correction.
• The study offers guidelines and tools for designing effective emulators for satellite data processing.
• Emulation successfully balances computational speed with the accuracy required for atmospheric correction applications.