The relevant range of scales for multi-scale contextual spatial modelling.

Area of Science:

• Environmental science
• Geostatistics
• Machine learning

Background:

• Spatial autocorrelation in environmental models often stems from missing covariate data.
• Incorporating covariates from multiple scales can mitigate this spatial autocorrelation.
• Identifying the relevant range of scales is crucial for accurate spatial modeling.

Purpose of the Study:

• To propose a systematic, data-driven method for determining the effective scale space in spatial modeling.
• To evaluate the impact of this effective scale space on machine learning model accuracy.
• To test the utility of variogram analysis for a priori approximation of the effective scale space.

Main Methods:

• Developed a data-driven method to objectively derive scale limits for spatial modeling.
• Applied machine learning techniques to assess modeling accuracy with the identified effective scale space.
• Investigated the use of variogram properties for predicting the effective scale space.

Main Results:

• Modeling using an effective scale space significantly improved machine learning model accuracy.
• A strong correlation was observed between variogram properties and the relevant range of scales.
• The variogram can effectively approximate the optimal scale space for contextual spatial modeling.

Conclusions:

• The proposed method enhances spatial modeling accuracy by defining an effective scale space.
• Variogram analysis serves as a valuable tool for a priori estimation of this scale space.
• Variograms are important for understanding structural dependencies in contextual spatial modeling and geostatistics.