Directional Change in Polygonal Distributions: Comparing human and computational directional relations in GIS data

Existing methods for calculating directional relations in polygons (i.e. the directional similarity model, the cone-based model, and the modified cone-based model) were compared to human perceptions of change through an online survey. The results from this survey provide the first empirical validation of computational approaches to calculating directional relations in polygonal spatial data. We have found that while the evaluated methods generally agreed with each other, they varied in their alignment with human perceptions of directional relations. Specifically, translation transformations of the target and reference polygons showed greatest discrepancy to human perceptions and across methods. The online survey was developed using Qualtrics Survey Software, and participants were recruited via online messaging on social media (i.e., Twitter) with hashtags related to geographic information science. In total sixty-one individuals responded to the survey. This survey consisted of nine questions. For the first question, participants indicated how many years they have worked with GIS and/or spatial data. For the remaining eight questions, participants ranked pictorial database scenes according to degrees of their match to query scenes. Each of these questions represented a test case that Goyal and Egenhofer (2001) used to empirically evaluate the directional similarity model; participants were randomly presented with four of these questions. The query scenes were created using ArcMap and contained a pair of reference and target polygons. The database scenes were generated by gradually changing the geometry of the target polygon within each query scene. The relations between the target and reference polygon varied by the type of movement, the scaling change of the polygon, and changes in rotation. The scenarios were varied in order to capture a representative range of variability in polygon movements and changes in real world data. The R statistical computing environment was used to determine the similarity value that corresponds with each database scene based on the directional similarity model, the cone-based model, and the modified cone-based model. Using the survey responses, the frequency of first, second, third, etc. ranks were calculated for each database scene. Weight variables were multiplied by the frequencies to create an overall rank based on participant responses. A rank of one was weighted as a five, a rank of two was weighted as a four, and so on. Spearman’s rank-order correlation was used to measure the strength and direction of association between the rank determined using the three models and the rank determined using participant responses.