Generalized covariate field (GCF): spatial-pattern and neighbourhood-distribution feature expansion improves geospatial prediction
- 1School of Design and the Built Environment, Curtin University
Abstract
Spatial prediction forms the basis for decision-making in Earth and environmental sciences, yet prediction accuracy is constrained by spatial heterogeneity and sparse and uneven sampling. Existing methods commonly respond by increasing algorithmic complexity while underusing the spatial information embedded in the covariates themselves. This study develops the generalized covariate field (GCF) to expand the covariate space by deriving two families of spatial features without increasing model complexity. First, spatial pattern features characterize local spatial structure, including spatial dependence, heterogeneity, geocomplexity, multiscale variation and local outlyingness, and are transferable across space. Second, neighbourhood distribution features summarize covariates through multiscale operators with geoscientific meaning. The resulting high-dimensional features are then selected through response-independent functional reduction and spatial-block stability selection. GCF is implemented in a vascular plant species-richness study across the Southwest Australian Floristic Region, and expands 12 covariates into 3276 candidate predictors. Results show that GCF improves prediction accuracy under both random and spatial-block validation with 8.2–21.1% higher spatial-block cross-validated R2 than existing spatial learning models. The findings support the primacy of data hypothesis that enriching the spatial representation of data, rather than only increasing model structural complexity, can substantially improve prediction and provide a basis for data-centric geospatial modelling.
Related
- From Equations to Geospatial Foundation models. What comes next for forest prediction?
- Trends in soil organic carbon prediction research using geospatial technologies and artificial intelligence: an exploratory bibliometric analysis (2005-2025)
- Towards generalized hydrological forecasting using transformer models for 120 h streamflow prediction
- Automatic Extraction of a Generalized Physics‐Based Feature From Time‐Lapse Active Source Ultrasonic Waveforms for Prediction of Shear Stress in Laboratory Friction Experiments
- Distinguishing run-up height from pressure distribution during avalanche impact on narrow obstacles: mechanisms and semi-empirical prediction from three-dimensional simulations
BibTeX
@article{song2026generalized,
title = {{Generalized covariate field (GCF): spatial-pattern and neighbourhood-distribution feature expansion improves geospatial prediction}},
author = {Yongze Song},
journal = {International Journal of Geographical Information Systems},
year = {2026},
month = sep,
doi = {10.1080/13658816.2026.2729719},
url = {https://doi.org/10.1080/13658816.2026.2729719},
note = {CC BY},
}
RIS
TY - JOUR
TI - Generalized covariate field (GCF): spatial-pattern and neighbourhood-distribution feature expansion improves geospatial prediction
AU - Yongze Song
PY - 2026
DA - 2026/09/10
JO - International Journal of Geographical Information Systems
DO - 10.1080/13658816.2026.2729719
UR - https://doi.org/10.1080/13658816.2026.2729719
AB - Spatial prediction forms the basis for decision-making in Earth and environmental sciences, yet prediction accuracy is constrained by spatial heterogeneity and sparse and uneven sampling. Existing methods commonly respond by increasing algorithmic complexity while underusing the spatial information embedded in the covariates themselves. This study develops the generalized covariate field (GCF) to expand the covariate space by deriving two families of spatial features without increasing model complexity. First, spatial pattern features characterize local spatial structure, including spatial dependence, heterogeneity, geocomplexity, multiscale variation and local outlyingness, and are transferable across space. Second, neighbourhood distribution features summarize covariates through multiscale operators with geoscientific meaning. The resulting high-dimensional features are then selected through response-independent functional reduction and spatial-block stability selection. GCF is implemented in a vascular plant species-richness study across the Southwest Australian Floristic Region, and expands 12 covariates into 3276 candidate predictors. Results show that GCF improves prediction accuracy under both random and spatial-block validation with 8.2–21.1% higher spatial-block cross-validated R2 than existing spatial learning models. The findings support the primacy of data hypothesis that enriching the spatial representation of data, rather than only increasing model structural complexity, can substantially improve prediction and provide a basis for data-centric geospatial modelling.
SN - 0269-3798
SN - 1362-3087
ER -