Using Machine Learning to Predict Column Concentrations and Retrieval Diagnostics of the TROPESS Atmospheric Composition Profiles

Advances in sensor technology have led to a substantial increase in the data output from satellite-borne remote sensing instrumentation. For example, NASAs Cross-track Infrared Sounder (CrIS) and Atmospheric Infrared Sounder (AIRS) provide millions of global, spectrally-resolved radiance observations every day. It is becoming increasingly challenging to process these large data sets and perform the subsequent composition profile retrievals for all observations. Indeed, NASAs TRopospheric Ozone and its Precursors from Earth System Sounding (TROPESS) project, which produces records of atmospheric constituents from multiple satellite and ground data through a common retrieval algorithm, can only process about 1% of the sampled CrIS observations. This talk presents efforts to process all the observed CrIS and AIRS data by applying machine learning techniques. In particular, we present staggered artificial neural networks (ANNs) that can reliably replicate the retrieved CrIS carbon monoxide and ammonia profiles, as well as important retrieval diagnostics such as the retrieval error and averaging kernels. Once trained, these ANNs can perform predictions for millions of CrIS radiance observations in minutes. This new data set not only covers the gaps in the global retrievals of composition fields, but also provide uncertainty and variability information on very small scales.