Gaussian Process and Deep Learning Atmospheric Correction

Atmospheric correction is the processes of converting radiance values measured at a spectral sensor to the reflectance values of the materials in a multispectral or hyperspectral image. This is an important step for detecting or identifying the materials present in the pixel spectra. We present two machine learning models for atmospheric correction trained and tested on 100,000 batches of 40 reflectance spectra converted to radiance using MODTRAN, so the machine learning model learns the radiative transfer physics from MODTRAN. We created a theoretically interpretable Bayesian Gaussian process model and a deep learning autoencoder treating the atmosphere as noise. We compare both methods for estimating gain in the correction model to process for estimating gain within the well-know QUAC method which assumes a constant mean endmember reflectance. Prediction of reflectance using the Gaussian process model outperforms the other methods in terms of both accuracy and reliability.