A Categorical Machine Learning Approach to Predicting Areas of Shallow Coastal Groundwater

Area of Science:

• Environmental science
• Geoscience
• Water resource management

Background:

• Coastal lowlands are increasingly vulnerable to sea-level rise and groundwater changes.
• Accurate groundwater depth estimation is crucial for coastal adaptation planning.
• Traditional regression methods often struggle with data imbalance and uncertainty quantification.

Purpose of the Study:

• To introduce a novel categorical modeling framework for groundwater depth estimation.
• To redefine groundwater depth estimation as a classification problem, incorporating uncertainty quantification.
• To provide a robust tool for evidence-based decision making in coastal management.

Main Methods:

• Developed a national-scale ensemble model at 100m resolution using the Random Forest algorithm.
• Trained the model on New Zealand's depth-to-water database with 37 principal component-derived predictor variables.
• Categorized groundwater occurrence into multiple depth thresholds (0.9-2.0m) to quantify prediction uncertainty via Type I and Type II errors.

Main Results:

• The ensemble model achieved strong performance with ROC-AUC values ranging from 0.823 to 0.962.
• Model accuracy demonstrated improvement with increasing groundwater depth.
• The categorical framework effectively addressed data imbalance and enhanced uncertainty quantification.

Conclusions:

• The categorical modeling approach offers improved uncertainty quantification compared to traditional regression methods.
• Probabilistic predictions enable customizable risk thresholds for stakeholders in coastal management.
• This framework bridges advanced modeling with practical adaptation planning for rising sea levels.