Ammonium nitrogen (NH4+-N) contamination in groundwater beneath chemical industrial parks exhibits extreme spatial heterogeneity, yet the comparative effectiveness of spatial prediction methods under such conditions remains poorly understood. At a chemical industrial park in a region of Shanxi Province, northern China, we analyzed 133 monitoring wells sampled across four campaigns (April–October 2024) at two aquifer depths. Global Moran’s I (0.040–0.118) and variogram nugget ratios (>75%) indicated weak spatial autocorrelation. Consequently, on the raw concentration scale, all six geostatistical methods yielded near-zero or negative leave-one-out cross-validation (LOO-CV) R2. Evaluated on the log10 scale, machine learning (ML) models achieved positive predictive skills, with Extreme Gradient Boosting (XGBoost) performing best (R2 ≈ 0.75). Three hybrid ML–kriging methods produced physically coherent plume surfaces while retaining their predictive skills; the April upper-layer result (R2 ≈ 0.67)—the only campaign without retained within-well information—best represents spatial generalization, whereas the higher later-campaign values (R2 > 0.97) are optimistic. Exceedance probability mapping based on XGBoost (area under the ROC curve, AUC = 0.959) revealed a persistent high-risk zone. Because the geostatistical and ML metrics span different response scales and validation schemes, their comparison is indicative rather than a direct ranking. Spatial autocorrelation diagnostics should precede method selection at point-source-dominated industrial sites.
Lü et al. (Tue,) studied this question.