Work

Spatial Support and Operator Structure in GSTAR Models: Administrative Regions versus Regular Grids for Wildfire Hotspot Data

Year: 2026
Type: preprint
Abstract: This study examines the impact of spatial support transformation on the structure and stability of the Generalized Space-Time Autoregressive (GSTAR) framework. We explore two types of spatial representation: irregular administrative regions and regular, high-density grid discretization. Administrative regions are typically defined by polygonal boundaries, and the heterogeneity of these polygons often leads to inconsistent neighborhood configurations and non-uniform spatial interactions. In contrast, regular grid discretization yields more structured, sparse spatial operators. Using spatio-temporal environmental data obtained from NASA FIRMS observations, GSTAR models were constructed using different spatial supports and evaluated based on spatial interaction structure, parameter stability, and forecasting characteristics. Beyond empirical comparison, this study investigates the structural properties of the resulting GSTAR operators, such as sparsity preservation, block-banded structure, and spectral properties within the regular grid topology. The results indicate that regular grid discretization is more efficient than irregular administrative aggregation at preserving local neighborhood interactions; this method produces spatial operators that are sparser, more localized, and spectrally controlled, which in turn leads to more stable and accurate spatio-temporal parameter estimation. On the other hand, forecasting accuracy proves to be highly dependent on the error measurement method: when compared at the original scale (without standardization), grid-based models yield lower absolute forecasting errors; However, once errors are standardized based on the variability of each spatial unit—an adjustment necessary given that spatially aggregated administrative data series tend to be structurally smoother—administrative support actually yields normalized errors that are comparable or even lower. We present both perspectives and argue that the primary structural advantage of grid support lies in operator sparsity and parameter stability, rather than in an absolute superiority in forecasting accuracy. These results demonstrate that spatial support is not merely a pre-processing step, but a structural element that fundamentally impacts the behavior of spatio-temporal operators.
Language: English
Open Access status: green
OpenAlex ID: W7225985436
Cites: 24
Cited by:
Related to: 10