Spatial coherence and input uncertainty in AutoML-tuned gradient-boosting predictions of depth to groundwater as an indicator of groundwater potential

Abstract Groundwater in the semi-arid Zagros foreland of northern Iraq is being withdrawn faster than it is replenished, and identifying where groundwater is most accessible is a precondition for managing that deficit. This study maps groundwater potential in the Karma Soor basin, Kurdistan Region, and compares three gradient-boosted ensemble learners, LightGBM, XGBoost and CatBoost, tuned under identical settings with the Train Using AutoML tool in ArcGIS Pro. Depth to groundwater served as the response variable, interpreted as an indicator of groundwater accessibility rather than of recharge, and twelve conditioning factors spanning the topographic, hydrological, geological, pedological and land-cover domains served as predictors. Under a random hold-out, XGBoost returned the lowest errors (MAE 5.21 m, RMSE 7.90 m, R² 0.905), but the bootstrap confidence intervals of the three models overlapped widely, and under four-fold spatial block cross-validation CatBoost generalized best (RMSE 6.89 m); the three models are therefore broadly comparable, with their ranking dependent on the validation design. Stream order dominated both impurity-based (38.62%) and SHAP importance, acting largely as a proxy for topographic position, whereas removing rainfall did not reduce accuracy. Perturbing the predicted surface by ± 5%, ± 10% and ± 20% across one hundred simulations and re-running the hot-spot analysis left 90.25%, 80.16% and 66.12% of the basin in its original category. Instability grew sub-linearly and concentrated at cluster margins, and it consisted mainly of shifts between adjacent confidence levels (e.g., from 95% to 90% hot spot) rather than jumps between hot-spot and cold-spot categories, while the cores of the high-potential channel corridors remained stable. By evaluating spatial generalization, spatial coherence and stability alongside conventional accuracy, the study shows that hold-out error statistics alone are an insufficient basis for selecting a groundwater potential model.

Authors

Publication Details

Journal
Scientific Reports
Published
2026-11-14
DOI
https://doi.org/10.1038/s41598-026-78986-1
Primary Topic
Groundwater and Watershed Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Spatial coherence and input uncertainty in AutoML-tuned gradient-boosting predictions of depth to groundwater as an indicator of groundwater potential

Zakariya Nafi Shehab, Raid Mahmood Faisal, Nadhir Abbas Al-Ansari
Scientific Reports
Groundwater and Watershed Analysis
article

Spatial coherence and input uncertainty in AutoML-tuned gradient-boosting predictions of depth to groundwater as an indicator of groundwater potential

Zakariya Nafi Shehab, Raid Mahmood Faisal, Nadhir Abbas Al-Ansari
article en

Abstract

Abstract Groundwater in the semi-arid Zagros foreland of northern Iraq is being withdrawn faster than it is replenished, and identifying where groundwater is most accessible is a precondition for managing that deficit. This study maps groundwater potential in the Karma Soor basin, Kurdistan Region, and compares three gradient-boosted ensemble learners, LightGBM, XGBoost and CatBoost, tuned under identical settings with the Train Using AutoML tool in ArcGIS Pro. Depth to groundwater served as the response variable, interpreted as an indicator of groundwater accessibility rather than of recharge, and twelve conditioning factors spanning the topographic, hydrological, geological, pedological and land-cover domains served as predictors. Under a random hold-out, XGBoost returned the lowest errors (MAE 5.21 m, RMSE 7.90 m, R² 0.905), but the bootstrap confidence intervals of the three models overlapped widely, and under four-fold spatial block cross-validation CatBoost generalized best (RMSE 6.89 m); the three models are therefore broadly comparable, with their ranking dependent on the validation design. Stream order dominated both impurity-based (38.62%) and SHAP importance, acting largely as a proxy for topographic position, whereas removing rainfall did not reduce accuracy. Perturbing the predicted surface by ± 5%, ± 10% and ± 20% across one hundred simulations and re-running the hot-spot analysis left 90.25%, 80.16% and 66.12% of the basin in its original category. Instability grew sub-linearly and concentrated at cluster margins, and it consisted mainly of shifts between adjacent confidence levels (e.g., from 95% to 90% hot spot) rather than jumps between hot-spot and cold-spot categories, while the cores of the high-potential channel corridors remained stable. By evaluating spatial generalization, spatial coherence and stability alongside conventional accuracy, the study shows that hold-out error statistics alone are an insufficient basis for selecting a groundwater potential model.

Scientific ReportsVol. 16(1)
Groundwater and Watershed Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon