Atmospheric Corrections for Earth Observation with Deep Learning

When measuring Earth’s optical reflectivity properties from space, it is necessary to correct for atmospheric effects like absorption, scattering, etc. Such atmospheric corrections, based on top-of-atmosphere images, are among the core challenges of satellite-based optical remote sensing. As an alternative to the use of elaborate physics-based models, we present an approach to implement atmospheric correction with a deep learning network. Sentinel-2 optical satellite image pairs were used: L1C processing level (uncorrected, top-of-atmosphere) and L2A processing level (bottom-of-atmosphere) that has been atmospherically corrected using a physics-based method. The deep learning network was trained to learn the corrections to produce L2A images from L1C inputs. The model was trained and tested on geographically and atmospherically diverse images obtained from all around the world, and achieved a mean R
²
-score of 0.943 and mean absolute error of 0.0112 on the test set. Wavelength bands with strong atmospheric absorption, such as the water-vapor band, pose some challenges which could lead to worse atmospheric correction. However, the deep learning model seems to handle this relatively adequately.