We introduce the ability to algorithmically analyze important steps of topological evolution in severe weather data. It uses the Spatiotemporal Relational Probability Tree (SRPT), a probability estimation tree for relational data that varies across space and time. The SRPT randomly samples from an almost infinite set of possible distinctions, or questions regarding the various spatial and temporal qualities of a dataset's objects, relations, and attributes, and builds a tree using the most significant ones. This paper shows how the addition of a full set of spatial ordering distinctions affects the predictive ability of the SRPT and allows it to produce more meteorologically significant observations in the domain of severe thunderstorm models.

The spatial ordering distinctions apply seven topological relations: disjoint, meet, overlap, covers, equal, and contains, to other distinctions within the algorithm. This joining of multiple distinctions within a single more general distinction creates a conjugate distinction. They can track how an object in the dataset changes in space, such as how the rear flank downdraft changes in relation to the hail core, by analyzing the spatial domain of the distinctions that encapsulate each of those objects. For instance, if the rear flank downdraft and the hail core share no points in space at one time, then they would be disjoint. Furthermore, the conjugate distinction framework allows for hybrid spatial ordering and temporal ordering distinctions that can detect changes across space and time simultaneously. In this example, the SRPT could determine if the rear flank downdraft and hail core were disjoint before the rear flank downdraft covered the hail core. A distinction of this type would be truly spatiotemporal.

The algorithm is validated on data extracted from a collection of model-generated severe thunderstorms. The storms were generated using the Advanced Regional Prediction System (ARPS), a mesoscale model designed for generating simulated storms. The SRPTs are evaluated on several datasets with a focus on simulated supercell thunderstorms. The performance of the algorithm is compared with and without the new spatial ordering distinctions.