On estimation of atmospheric scattering characteristics from spectral measurements of solar radiation using machine learning algorithms

Abstract

Determination of the content of trace gases (such as, for example, NO2, formaldehyde) by DOAS (differential optical absorption spectroscopy) method in the lower troposphere can be difficult with significant scattering of light on clouds and aerosol. Often, the parameters of cloudiness and aerosol are unknown for specific DOAS measurements, and, therefore, the estimation of these parameters directly from the DOAS analysis data is an approach that could increase the final measurement accuracy of trace gases. In this work, we consider the problems of retrieving such characteristics as: cloud bottom height, cloud optical depth, aerosol optical depth, F-factor (a factor reciprocal to air mass factor) from the input data obtained during DOAS analysis. To do this, we trained and compared two machine learning (ML) models - neural network and random forest. Both ML algorithms solve the regression problem; data obtained by numerical computation by linearized radiative transfer model were used as a learning dataset. The dependence of the error on the test dataset depending on the regularizing parameters was investigated for the neural network. Retrieval errors of aerosol and cloud characteristics were preliminary estimated.