Simulation-Based Error Assessment for AIRS Near-Surface Temperature Retrievals: A Machine Learning Approach

A remote sensing retrieval combines observed radiances with a physical model of the interaction of radiation with the atmosphere and surface to estimate geophysical quantities of interest. The Atmospheric Infrared Sounder (AIRS) uses a multi-stage method to estimate vertical profiles of temperature and water vapor in clear and cloudy conditions. The statistical properties of the retrieval algorithm can be diagnosed both through validation as well as through controlled simulation experiments of the observing system. This work illustrates a Monte Carlo experiment framework for the AIRS retrieval that includes a probabilistic model for the joint distribution of the atmospheric and surface state. We present a series of retrieval simulation experiments for multiple cloud regimes corresponding to the Marine ARM GPCI Investigation of Clouds (MAGIC) campaign. These experiments are used to train random forest regression models for predicting near-surface temperature retrieval error as a function of retrieved quantities, and the method is tested against the MAGIC campaign validation data.