A machine learning-based method to account for 3D Short-Wave radiative effects in 1D satellite observation operators. (November 2021)

• Record Type: Journal Article
• Authors: Zhou, Yongbo; Liu, Yubao; Liu, Chao
• Is Part Of: Journal of quantitative spectroscopy & radiative transfer. Volume 275(2021)
• Publication Date: 2021-11
• Subjects: Observation operator -- 3D radiative effects -- Machine learning; Spectrum analysis -- Periodicals; Radiation -- Periodicals; Analyse spectrale -- Périodiques; Rayonnement -- Périodiques; Radiation; Spectrum analysis; Periodicals; 543.0858
• DOI: 10.1016/j.jqsrt.2021.107891
• Languages: English
• ISSNs: 0022-4073

Abstract

Highlights: We show a machine learning method to correct 1D short-wave observation operators. The method correct cloud 3D radiative effects accurately and fast. The method is applicable to current 1D observation operators. Abstract: Data assimilation of Short-Wave (SW, ≤ 4μm) satellite radiance in cloudy regions could potentially improve cloud forecasts. Operational SW observation operators neglect cloud 3D radiative effects, which may cause nonnegligible errors in some circumstances, for example, with large solar zenith angles or for broken clouds. This study introduced a machine learning (ML)-based method to include some cloud 3D radiative effects in a 1D observation operator based on liquid water cloud simulations by the Weather Research and Forecasting model and radiative transfer simulations by the Spherical Harmonics Discrete Ordinate Method. The inputs of the ML correction method include the reflectance simulated by a 1D observation operator at the targeting column, cloud top height, cloud water content, and effective particle size at its adjacent columns, and other ancillary information (sun-viewing geometries, water vapor conditions, etc.). Preliminary results were presented for the channel 3 (0.75 ∼ 0.90µm) of the Advanced Geostationary Radiation Imager (AGRI) onboard FY-4A. Compared with 1D simulations, the ML correction method could reduce the biases from 6.9% ∼ 11.7% to 2.0% ∼ 4.7% and the root mean square errors from 22.3% ∼ 34.8% to 15.4% ∼ 27.4%. The elapsed CPU time for the ML correction method is approximately the same or two times that of the 1D observation operators, depending on whether atmospheric gases are included or not. In general, the ML correction method is computationally more efficient than traditional 3D radiative transfer solver, and could be used to correct current 1D observation operators' simulations.