Many real world domains, such as severe weather events, are inherently spatiotemporal in nature. Each year severe weather induced by thunderstorms causes property damage, injury, and loss of life. Convectively induced turbulence is a hazard to airlines, which at best requires rerouting flight paths, but can lead to significant delays and even structural damage to the aircraft and loss of life. Tornados are possibly the most impressive and destructive potential product of thunderstorms. Domains such as severe weather require a system that is capable of representing and reasoning about complex spatiotemporal data. The dynamics of attributes and relationships varying both spatially and temporally provides a unique set of challenges.
Spatiotemporal Relational Probability Trees (SRPTs) are a type of decision-tree that reason with complex spatiotemporal relational data. High level objects and their relationships are extracted from the raw low level dataset. This allows the SRPTs to reason in more abstract terms and to use relationships that are critical to understanding how things interact. Combining SRPTs into random forests creates a Spatiotemporal Relational Random Forest (SRRF), which is capable of capturing more varied and complex concepts than single trees.
This thesis introduces significant enhancements to SRPT that increases it's ability to reason about spatiotemporal data. Often, high-level objects come from scalar- or vector-valued two- or three-dimensional temporal regions called fields. Previously these fields were discarded after the generation of high-level objects. We add the ability to reason about both the objects and the fields within the objects. These fields allow us to add the ability to ask question about the gradient, divergence, and curl of those fields. We also add the ability to recognize the shape of fields, allowing for questions regarding change of shape and orientation. Lastly, we add the ability to reference a single object within the data, and simple boolean operations for combining two questions. These additions are validated using SRRFs on a several real-world.
The SRRF algorithm learns robust classifiers on each of the domains, either outperforming the SRRF without fields or performing equally well. Analysis of the forests produced showed that features the SRRF algorithm used were consistent with meteorological theories. We show that the addition of fields can be a valuable resource to the SRRF algorithm for spatiotemporal analysis.
Troutman, Nathaniel. (2010). Enhanced Spatiotemporal Relational Probability Trees and Forests. Master's Thesis, School of Computer Science, University of Oklahoma.
SRPT and SRRF Code
Created by amcgovern [at] ou.edu.
Last modified June 12, 2017 12:57 PM