Applying Machine Learning Techniques to Retrieve and Improve ATMS-Based Temperature, Water Vapor, and Precipitation Estimates

This study will apply machine learning (ML) techniques to improve and retrieve atmospheric parameters based on The Advanced Technology Microwave Sounder (ATMS) data including radiance observations or physical retrievals. The ATMS observes the Earth in the microwave portion of the electromagnetic spectrum through 22 channels which allows it to see through clouds and provides information about the physical proprieties of the atmosphere, such as temperature and hydrometeors. The instrument is currently onboard SNPP, NOAA-20 and NOAA-21. The NOAA Microwave Integrated Retrieval System (MiRS) is a physical based algorithm and retrieves vertical profiles of water vapor, cloud liquid water, graupel water, rain water, and temperature, and is run operationally on all available ATMS data.

One of the advantages of ML techniques is the ability to solve non-linear problems solely based on input data without the need to explicitly model complex physical systems. In addition, once the model is trained, the application is computationally efficient and generally significantly faster than a physics-based model. We will investigate the potential of ML techniques to improve ATMS-based estimates of atmospheric parameters, focused in two areas.

First, we will address the improvement of atmospheric temperature and water vapor retrievals. The collocated European Centre for Medium-Range Weather Forecasts (ECMWF) data will be used as training target. A deep neural network architecture will be used with several fully connected layers and on the order of 100 neurons on each layer. The MiRS system will be included in this analysis as both a comparison data set, and as a potential input to a ML model.

Second, we will investigate the use of ML methods to improve precipitation estimates from MiRS ATMS retrievals. Our previous study applied a U-Net architecture and used MiRS retrieved hydrometers as inputs to predict surface precipitation rate. The training target data were hourly precipitation rates from the operational Multi-Radar/Multi-Sensor System (MRMS). Results showed significant improvement over MiRS retrievals over the Continental U.S. (CONUS). This study will generalize the U-Net model to Asia and Europe. The ground truth data are the AERA5-Asia (Asian Precipitation-Highly Resolved Observational Data Integration Toward Evaluation of Water Resources European Centre for Medium-Range Weather Forecasts Reanalysis 5-Land) over Asia, and the European Climatological high-resolution gauge-adjusted radar precipitation dataset (EURADCLIM) over Europe.