Answers are grounded in this paper and cite their sources.
Random cell-level validation can overestimate flood-susceptibility discrimination when nearby cells share inundation features. We assessed this effect in Yeoju, Republic of Korea, using 2,000 flood polygons from 2006 and 2013, 82 leakage-control groups, 12 static factors and nine models. Five repeats of five-fold grouped spatial validation yielded lower AUCs than random hold-outs, with differences of 0.017–0.136. GAT achieved the highest mean out-of-fold AUC (0.951 ± 0.012 SD), although this ranking remains conditional on the study region and implementations. Sensitivity experiments examined fold count, spatial grouping, buffering, predictor encoding and graph construction. Additional metrics assessed classification, calibration and agreement with inventory labels. Associations between prediction dispersion and score–label discrepancy depended on class composition: RGCN and random forest showed negative correlations within mapped flood cells, whereas XGBoost showed a negative whole-raster correlation. These findings demonstrate that validation design, predictor representation and class composition influence model assessment. They support grouped spatial evaluation for screening spatially withheld areas within Yeoju and class-specific checks before interpreting dispersion as an instability warning.
No takes yet. Share an insight, caveat, or question.
Han et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: