Random forest as a generic framework for predictive modeling of spatial and spatio-temporal variables.

Area of Science:

• Geospatial analysis
• Machine learning
• Environmental modeling

Background:

• Traditional machine learning models often overlook spatial autocorrelation, potentially leading to biased predictions.
• Ignoring geographical location in spatial predictions can result in suboptimal model performance and unreliable outputs.

Purpose of the Study:

• To present a novel Random Forest for Spatial Predictions (RFsp) framework that integrates geographical proximity into the modeling process.
• To evaluate the performance of RFsp against state-of-the-art kriging techniques for various spatial and spatio-temporal prediction tasks.

Main Methods:

• Developed the RFsp framework using buffer distances from observation points as explanatory variables.
• Applied RFsp to textbook datasets for numeric, binary, categorical, multivariate, and spatio-temporal predictions.
• Compared RFsp performance with kriging using fivefold cross-validation with refitting.

Main Results:

• RFsp achieved prediction accuracy and unbiasedness comparable to kriging methods.
• RFsp demonstrated flexibility in handling diverse covariate types and required fewer statistical assumptions.
• RFsp offers potential for more informative error maps, especially for multivariate spatial prediction models.

Conclusions:

• RFsp is a viable and flexible alternative to kriging for spatial predictions, particularly in geoscience applications.
• The framework's success hinges on high-quality training data, robust spatial sampling, and rigorous model validation.
• While computationally intensive with large datasets, RFsp provides a powerful tool for spatial and spatio-temporal analysis.