Authors: Henry Hui Luan, University of Oregon, Antoine Nzeyimana*, University of Oregon
Topics: Geographic Information Science and Systems, Spatial Analysis & Modeling, Cyberinfrastructure
Keywords: spatial data models, representation learning, deep learning, GeoAI, machine learning, spatial analysis
Session Type: Poster
Start / End Time: 8:00 AM / 9:40 AM
Room: Lincoln 2, Marriott, Exhibition Level
Presentation File: Download
Representation models of spatial data influence the kinds of analyses we can perform and thus affects the conclusions we can make about the data. Choosing between different representation models may also lead to different accuracies in our analysis results. Data representation also impacts the computational performance (memory efficiency and run time) of our analysis algorithms, an aspect which is critical for the processing big spatial data. Various geo-spatial problem domains have established multiple ways of representing geographic data. For instance, the raster data model which represents space as a regular grid is used to model continuous fields on the earth surface and it is commonly used in areas such as remote sensing and climate modeling. Other spatial data models exist to represent irregular and more complex data indexed by geo-spatial coordinates. These include vector data models, irregular tessellation and spatial network models. This study aims to understand the impact of spatial data representation in geo-spatial problem solving. We want to understand for instance which information is lost, retained or enhanced when we choose to work with one representation over the other. The success of representation learning in machine learning models also makes us wonder how to learn a representation that preserves spatial and spatio-temporal properties and concepts of nearness and auto-correlation and the impact of the learned representation on the uncertainty inherent in geographic data.