Authors: Bin Li*, Central Michigan University
Topics: Quantitative Methods, Spatial Analysis & Modeling, Geographic Information Science and Systems
Keywords: regression, raster, uncertainty, eigenvector spatial filtering
Session Type: Paper
Start / End Time: 5:20 PM / 7:00 PM
Room: Galvez, , Marriott, 5th Floor
Presentation File: No File Uploaded
Spatial autocorrelation results in variance inflation in parameter estimations, leading to greater uncertainty. Such an effect is particularly prominent in raster-based regression modeling where a high level of positive spatial autocorrelation is often present. While geostatistical approaches may remedy the situation, they have limitations, particularly when dealing with non-Gaussian variables. Recent developments in Eigenvector Spatial Filtering (ESF) offer a potential alternative. In the core of the ESF method is the representation of spatial information as a linear combination of a selective set of eigenvectors of the spatial weights matrix. By incorporating the ESF in regression models, useful information can be filtered from the residual and moved to the model, which enhances the quality of the parameter estimates. While numerous applications have demonstrated the capability of ESF-based regression modeling, they are largely limited to the vector data model. Raster-based regression modeling with ESF has been a challenge due to the computation intensity involved. In this paper, we report a potential solution using a local strategy where an ESF is composed for each cell and added to the global model. The experiment is based on a set of simulated data and several real-world applications. Standard model assessment methods are used to evaluate the approach. Initial experiment showed noticeable improvement in narrowing uncertainty of parameter estimations.