Applying machine learning to improve the near-real-time products of the Aura Microwave Limb Sounder

A new algorithm to derive near-real-time (NRT) data products for the Aura Microwave Limb Sounder (MLS) is presented. The old approach was based on a simplified optimal estimation retrieval algorithm (OE-NRT) to reduce computational demands and latency. This paper describes the setup, training, and evaluation of a redesigned approach based on artificial neural networks (ANN-NRT), which is trained on >17 years of MLS radiance observations and composition profile retrievals. Comparisons of joint histograms and performance metrics derived between the two NRT results and the operational MLS products demonstrate a noticeable statistical improvement from ANN-NRT. This new approach results in higher correlation coefficients, in addition to lower root-mean-square deviations and biases at almost all retrieval levels compared to OE-NRT. The exceptions are pressure levels with concentrations close to 0 ppbv (parts per billion by volume), where the ANN models fail to establish a functional relationship and tend to predict 0. Depending on the application, this behavior might be advantageous. While the developed models can take advantage of the extended MLS data record, this study demonstrates that training ANN-NRT on just a single year of MLS observations is sufficient to improve upon OE-NRT. This confirms the potential of applying machine learning to the NRT efforts of other current and future mission concepts.

Automated summary

A new algorithm based on artificial neural networks (ANN-NRT) is presented to derive near-real-time data products for the Aura Microwave Limb Sounder (MLS). Statistical improvements are observed compared to the old approach, with higher correlation coefficients, lower root-mean-square deviations, and biases. This study confirms the potential of applying machine learning to near-real-time efforts of other mission concepts.