Deep Learning Based Emulation of Radiative Transfer Code for Atmospheric Correction of Satellite Images

Atmospheric correction of satellite images in the remote sensing area is one of the main pre-processing techniques since the better the effects of the atmosphere are eliminated in these images, the more and better features can be extracted later. Atmospheric correction consists of obtaining the reflectance of the surface (whether the surface of the earth or the surface of water, also known as “water leaving reflectance”), taking into account the effect of the atmosphere on the electromagnetic spectrum both by scattering and by the absorption of sunlight. This work proposes the use of Deep Learning models (DL) to obtain this reflectance of the earth's surface using synthetic data generated for different atmospheric conditions (aerosols, water vapor, ozone) and geometric conditions during the satellite flight, using simulations in a Radiative Transfer Code (6S) and data from the image itself that allow training a DL model to make this atmospheric correction. The models evaluated in this work are a Multilayer Perceptron (MLP), a 1D Convolutional Neural Network (1D-CNN), a Multilayer Perceptron combined with “Numerical Embedding” (MLP-NE) and an architecture based on the “Vision Transformer”. (ViT)” using these atmospheric parameters to feed the DL model. The results obtained by the models are compared with the numerical simulation, concluding that the last two models in particular have quite accurate performance, with the advantage that the inference time of the DL model reduces the calculation time with respect to the simulation with close precision in some cases.