Capturing relationships between coherent structures and convectively-induced turbulence using Spatiotemporal Relational Random Forests

Spatiotemporal relational random forests (SRRFs) provide a novel extension of the traditional random forest technique by using spatiotemporal relational probability trees (SRPTs) to comprise the forest ensemble. SRPTs are well-suited to identifying critical spatial, temporal, and spatiotemporal features of a domain from a "training" data set. This paper details how SRRFs are applied to the challenging problem of convectively-induced turbulence (CIT) prediction, with the goal of improving probabilistic CIT nowcasts for NextGen.

The CIT training data are represented as as attributed graph with coherent observed structures or objects such as 'aircraft' (representing a quantitative turbulence measurement location), 'area of convection' or 'area of hail' as vertices (each with several constituent attributes), and relationships such as 'nearby' or 'contains' as edges. The SRRF technique is used to determine which relationships are significant for predicting CIT intensity, and to create an empirical model in which the trees' output probabilities are combined and calibrated to produce a probabilistic nowcast of moderate-or-greater turbulence. The SRRF model is run and evaluated on several case studies. Conclusions are drawn regarding the relationships between observed storm features and NWP-derived environmental information and the formation of CIT. CIT prediction accuracy and computational cost tradeoffs are discussed with a view toward utilizing SRRFs in future operational CIT nowcasting products.