@phdthesis{01M0YH8Z7DBZYZ2HK6TZJ8JAWN,
  abstract     = {{Geographic phenomena are increasingly observed as sequences, and the questions posed of those sequences are questions of similarity: which spatial units behave alike, in what ways, and where their resemblances and differences are located. The standard pipeline for unsupervised clustering, however, was built around the assumption that observations live in a Euclidean feature space and that distances behave as metrics. Elastic similarity measures such as Dynamic Time Warping (DTW), developed to compare sequences whose features are shifted along the time axis, do not satisfy that assumption. Their adoption in geographic research has accelerated over the past decades, but the methodological implications of working with a non-metric similarity measure have remained unevenly addressed: which adaptations the clustering pipeline requires at which junctures, and which it does not, is not settled.
This dissertation develops a similarity-based framework for unsupervised clustering of geographic time series in which DTW plays a central role. Three interconnected methodological problems organize the work: the choice of similarity measure, namely when and why DTW is preferable to a simpler pointwise alternative; the bridge from a non-metric dissimilarity matrix to a clustering algorithm, namely when dimensionality reduction is necessary as an intermediate representation and when it is not; and validation, namely how to assess cluster quality when standard internal indices lose their formal footing. The framework is geographic rather than generic: it foregrounds the properties of geographic time series that condition the analysis at every stage, including spatiotemporal autocorrelation, the modifiable areal and temporal unit problems, and the changing spatial supports on which long observation windows are recorded.
Three contrasting empirical settings exercise the pipeline: a comparison of Euclidean distance matching and DTW on weekly COVID-19 incidence curves across European NUTS 2 regions; a typology of six-year traffic saturation on Flemish freeway segments; and a regime analysis of landscape transformation in Belgium based on hemeroby trajectories from CORINE Land Cover inventories, yielding the first national-scale characterization of landscape change framed explicitly in terms of human influence intensity.
The dissertation delivers two complementary contributions. The first is methodological: a conditional characterization of when each pipeline choice is warranted, together with a multi-criteria validation practice in which internal indices, stability analysis, spatial diagnostics, and cartographic interpretation function as complementary sources of evidence whose disagreements are themselves diagnostic. The second is domain-specific: typologies and regime analyses in spatial epidemiology, transport geography, and landscape ecology that demonstrate the framework's capacity to produce defensible geographic conclusions in domains where the standard clustering toolkit does not transfer without adaptation.}},
  author       = {{De Sloover, Lars}},
  language     = {{eng}},
  pages        = {{245}},
  publisher    = {{Ghent University. Faculty of Sciences}},
  school       = {{Ghent University}},
  title        = {{On the warp and woof of pattern discovery in geographic time series : a methodological framework for unsupervised learning with dynamic time warping in a geospatial context}},
  year         = {{2026}},
}