Abstract
Abstract
Groundwater reconnaissance maps can prioritize field verification where observations are sparse, but spatial clustering and unverified background samples can make small model differences appear more certain than they are. This study compared four spatial model structures for relative well-occurrence favorability in Cheongju, South Korea: geographical random forest (GRF), k-nearest-neighbor label-lag logistic regression (kNN-lag), coordinate-augmented extreme gradient boosting (Coord-XGBoost), and a spatial Gaussian process classifier (SGP). The analytical sample comprised 207 georeferenced positive well locations and 207 sampled background points. Because a common pumping-rate or specific-capacity threshold was unavailable, outputs were interpreted as conditional rankings rather than aquifer-yield probabilities. Thirteen conditioning factors on a 10 m grid were evaluated using a shared five-fold spatial-block partition and a secondary stratified hold-out set. Mean spatial-partition area under the receiver operating characteristic curve (AUC) ranged from 0.850 to 0.871, and hold-out AUC ranged from 0.858 to 0.875. Pairwise DeLong tests detected no hold-out differences (minimum p = 0.296), and estimated power to detect an AUC difference of 0.03 was 0.312. Fixed-configuration factors-only variants differed from full models by 0.000-0.010 in mean AUC, whereas spatial-only variants yielded 0.538–0.640. Model-specific attribution and fold-refit dispersion provided complementary information about recurrent factor associations and within-model score stability. The outputs support staged groundwater reconnaissance and monitoring, not estimates of yield, sustainable abstraction, groundwater quality, or drilling success.