A Deep Learning Perspective to Atmospheric Correction of Satellite Images

Atmospheric correction (AC) is the process of retrieving correct surface reflectance (SR) values from the top of the atmosphere (TOA) radiance values by removing the effect of the atmosphere from the electromagnetic radiation reflected from the earth. Many physics-based approaches perform AC. However, these approaches, due to the complex relationship among the factors involved in AC, are compute intensive and rely on precomputed lookup tables that are based on approximations of the true relationships among these factors. The rapid growth in computational power, advancements in remote sensing technology, availability of vast amounts of satellite imaging data, coupled with advances in tools, techniques and algorithms in machine and statistical learning, has resulted in an opportunity to employ Deep Learning (DL) based approaches for providing effective solutions for AC. In this paper, we explore the potential of deep learning for AC. We categorize and review three approaches for AC: DL assisted physics-based approach, physics aware DL approach and physics agnostic DL approach. The paper overviews each of these approaches by providing the rationale behind them, and discusses key open issues and highlights possible solutions in different contexts for AC using DL.