Development and Evaluation of Geostatistical Methods for Non-Euclidean-Based Spatial Covariance Matrices

Area of Science:

• Geostatistics
• Spatial Statistics
• Environmental Modeling

Background:

• Traditional geostatistical modeling uses Euclidean distances, which are unsuitable for complex spatial variations.
• Non-Euclidean distances are crucial in settings like bodies of water, but current methods lack validity.
• Existing solutions like Multi-Dimensional Scaling (MDS) for non-Euclidean distances introduce bias in covariance matrices.

Purpose of the Study:

• To propose and validate a novel geostatistical method for handling non-Euclidean distances.
• To ensure the positive-definiteness of spatial covariance matrices with non-Euclidean metrics.
• To improve the accuracy and reliability of geostatistical predictions in complex spatial settings.

Main Methods:

• A new method to re-estimate spatial covariance structures based on non-Euclidean distances.
• Comparison with standard Euclidean distance methods.
• Evaluation against a Multi-Dimensional Scaling (MDS) based approach using cross-validation.
• Testing on both simulated and real-world datasets.

Main Results:

• The proposed method ensures valid spatial covariance matrices for non-Euclidean distances.
• Existing MDS methods exhibit significant bias in prediction variance.
• The new method demonstrates a superior balance between prediction accuracy and variance.
• The proposed approach outperforms existing methods on both simulated and real-world data.

Conclusions:

• The developed method provides improved geostatistical predictions when non-Euclidean distances are necessary.
• It ensures the mathematical validity of spatial covariance structures.
• This offers a more reliable alternative for spatial modeling in complex environments.