Content area
Abstract
When working with spatiotemporal data, it is often desirable for experts to associate class labels with a space, across time, according to observations within that domain (for instance, determining land use/land cover). This is a time-consuming process for experts, so it is beneficial to enlist computational assistance to accelerate the process. Many approaches to this problem treat it as an extension of the general semi-supervised learning problem: given a set of feature vectors, some of which are labeled, generate class labels for the unlabeled ones. However, in spatiotemporal domains, data may not be available in the form of full feature vectors: each element in these nominal vectors may be sampled heterogeneously – one may come in the form of rasters, another in the form of time series, another potentially as unstructured instantial observations in space and time.
The Class Label Conflation Problem is a solution to this framing issue that accommodates the realities of data collection in a spatiotemporal domain by considering each source of data to be providing partial evidence for the class label at a given location and time. The evidences from these different sources are combined into a mass of evidence that can be conflated into the class label of the point in question.
This dissertation marks the first definition and examination of the Class Label Conflation Problem in the literature. We provide and analyze algorithms for several versions of the Class Label Conflation Problem: as constrained to raster sources, as constrained to point-based time series sources, and with a final version using alternate distance measures to find more accurate predictors. Experimental results on these versions demonstrate the effectiveness of algorithms based on this problem formulation.
