3A.3 Improving Satellite Observation Utilization for Model Initialization Through the Use of Supervised Machine Learning

Friday, 28 July 2017: 2:00 PM
Constellation E (Hyatt Regency Baltimore)
Yu-Ju Lee, NOAA/ESRL/GSD (CIRES), Boulder, CO; and C. Bonfanti and J. Q. Stewart

Within the National Oceanic and Atmospheric Administration (NOAA), an important scientific goal to is provide timely and accurate weather prediction. A significant part of weather prediction is related to the accuracy of model initialization through the assimilation of a data from a variety of earth system observations. In areas where in-situ observations are difficult, such as large spans of ocean or remote areas of land surface, data from satellite observations are used to help complete the initial conditions. At present, the volume of satellite data collected daily exceeds the ability of forecast models to make use of it. The satellite data assimilation process is computationally expensive and data are often reduced in resolution to allow timely incorporation into the forecast. This problem is only exacerbated by the recent launch of Geostationary Operational Environmental Satellite (GOES)-16 satellite providing several order of magnitude increase in data volume.

Through our research, we are using supervised machine learning to find and mark regions of interest (ROI) to lead to targeted extraction of observations from satellite observation systems to help improve the overall utilization and the impact of satellite data on weather forecasts. Our preliminary efforts to identify the ROI’s are focused on two areas. One is to find important features using the analysis data from the National Center for Environmental Prediction (NCEP) Global Forecast System (GFS) weather model. Another is through the use of feature identification on Water Vapor and other satellite imagery.

This presentation will provide an introduction to our approach to machine learning to identify these ROI’s, discussion on data processing and feature engineering to extract important observation features, the relationships between grid size and accuracy rate to detect weather phenomena, as well as ongoing and future activities related to this project.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner