Gaussian Process and Deep Learning Atmospheric Correction

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