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

Coastal lowlands are increasingly vulnerable to threats from sea-level and associated groundwater rise. This study introduces a categorical modeling framework that redefines groundwater depth estimation as a classification problem rather than a continuous prediction task. By dividing groundwater occurrence into multiple depth thresholds (0.9-2.0 m), the approach explicitly quantifies prediction uncertainty through Type I (false positive) and Type II (false negative) errors. A national-scale ensemble model developed at 100 m resolution using the Random Forest algorithm was trained on New Zealand's comprehensive depth-to-water database. Thirty-seven predictor variables, derived via PCA (97.5% variance retained) from 199 base predictors, were incorporated to capture the complex interactions influencing groundwater depth. The model demonstrates strong performance, with ROC-AUC values ranging from 0.823 to 0.962, and accuracy improves with increasing depth. This categorical framework addresses challenges associated with data imbalance and enhances uncertainty quantification compared to traditional regression methods. Probabilistic predictions allow stakeholders to set customizable risk thresholds and manage acceptable error levels based on specific coastal management contexts. By bridging the gap between advanced numerical modeling and practical adaptation planning, the approach provides a robust tool for evidence-based decision making in the face of rising sea levels.