Atmospheric Thermodynamic Profiling with Neural Networks

This paper reviews recent progress in neural-network-based atmospheric parameter retrieval algorithms. Satellite-based observations of the atmosphere are related to atmospheric and surface variables by the radiative transfer equation. Direct inversion of the radiative transfer equation is often intractable as the underlying atmospheric phenomenology is usually neither linearly related to the observations nor Gaussian. Neural networks provide a computationally efficient alternative with their relatively simple structure. However, effective use of neural network techniques in problems of this nature requires careful preprocessing of the input data, and this preprocessing is often highly dependent on the remote sensing scenario under consideration. In this paper, we present techniques for preprocessing neural network input data in the spectral, spatial, and temporal domains. The utility of these techniques is demonstrated with several applications, including the estimation of the atmospheric temperature profile, water vapor profile, and precipitation rate using satellite-based passive microwave and infrared measurements.