Using Machine Learning to Superob Observations for Use in Aerosol Data Assimilation

Abstract:

When faced with billions of data points to assimilate into an atmospheric model, thinning or superobbing the observations becomes necessary. The number of observations must be reduced due to computational constraints, but the selection process must be completed in a reasonable time and the result must produce a subset which best represents the state of the atmosphere. By using machine learning, we can develop an intelligent superobbing process which better meets these goals, compared to previous static superobbing or thinning methods. The National Centers for Environmental Prediction (NCEP) is making progress towards improved air quality as well as subseasonal to seasonal (S2S) prediction by developing advanced global aerosol modeling and analysis capabilities. Here we present results on initial work towards superobbing Aerosol Optical Depth (AOD) data from the Visible Infrared Imaging Radiometer Suite (VIIRS), to support future atmospheric composition data assimilation capabilities at NCEP. To create training data, we selected small regions (1° latitude by 1.25° longitude) of VIIRS data to serve as an input to the machine learning model, and as an output to compare against we chose AOD values provided by NASA’s retrospective MERRA-2 reanalysis. We found a simple two-layer, fully connected neural network approximates the MERRA-2 dataset reasonably well and preserves larger-scale features, while reducing the number of observations by several orders of magnitude.