Denoising Stratospheric Nadir Sounder Observations using a Machine Learning Technique for Gravity Wave Detection

Satellite observations of the atmosphere are often extremely noisy due to both hardware limitations and the inherent complexity of retrieving and making measurements of the atmosphere. Gravity waves, which are low amplitude signals present in the atmosphere, are hard to resolve in this data due to their relatively low amplitude and small spatial extent. As a result, noise becomes a limiting factor when trying to identify and characterise them in real observed data.

Current methods to address this problem often lean upon smoothing approaches; however, such approaches suppress small scale signals and reduce measured amplitude and momentum fluxes significantly. This impedes the process in developing the next generation of models where these waves must be resolved accurately.

A novel supervised machine learning approach is introduced which is able to accurately remove small scale noise features from nadir observations of gravity waves. This model was trained on synthetic observations derived from high resolution DYAMOND model runs.  This is then applied to 22 years of NASA AIRS data and 12 years of MetOp IASI data and used to produce a new gravity wave climatology to better access small amplitude gravity waves.