This paper describes a technique for fusing satellite observations and derived features with data from numerical weather prediction (NWP) models and other readily-available sources to create nowcast products suitable for operational use. The resulting products are more accurate than either satellite-based detections or NWP model forecasts alone, since they combine a comprehensive but imperfect estimate of the state of the environment with observations that capture limited characteristics of the phenomena but in a more timely and accurate way. The authors argue that this sort of data fusion approach may be worthwhile for other applications in which satellite data is used to help address the needs of a particular user community.

The data fusion method is based on building and calibrating ensembles of decision trees, known as random forests, to aid the selection of relevant quantities and create a statistical predictive model that may be used to generate nowcasts. Two NASA-funded efforts illustrate the approach. The first makes use of Global Forecast System model forecasts and geostationary satellite data, including algorithms developed under NASA and GOES-R funding for detecting features related to turbulence including convective overshooting tops, tropopause folds and downslope winds. A statistical model is developed to produce gridded turbulence nowcasts intended for eventual use by forecasters at the World Area Forecast Centers with the goal of enhancing aviation safety and efficiency. The second explores how Rapid Update Cycle model and satellite data, including the SATCAST convective initiation algorithm, may be used to improve short-range forecasting of storms over the Gulf of Mexico. In both cases, a random forest empirical model is built to associate antecedent NWP model forecast quantities and satellite observations with subsequent truth measurements (turbulence from aircraft observations or convection from radar measurements, respectively). Elevated turbulence and convective initiation events are rare, requiring them to be oversampled in the training dataset to ensure the empirical model's sensitivity. The model must then be calibrated to produce accurate results. The resulting nowcasts are evaluated based on case studies and statistical verification.