Zherui Xu

University of Delaware

Subject Areas: hydrology

 Recent Activity

ABSTRACT:

Groundwater pumping in agricultural aquifers reflects both hydroclimatic conditions and anthropogenic controls, but their relative importance can change over time. We analyzed annual pumping in the High Plains Aquifer Hydrologic Observatory Area (HPAHOA) from 1960 to 2020 using Deep Temporal Clustering, Bayesian changepoint detection, regime-aware XGBoost, and SHAP attribution. Approximately 13,000 cell-level pumping trajectories separated into three behavioral clusters, with asynchronous changepoints in 1974, 2000, and 2005. Models trained only on pre-changepoint data transferred poorly to post-changepoint periods, while models including post-changepoint information showed the greatest improvement in Cluster 3, indicating regime-dependent predictor–response relationships. SHAP attribution showed that dominant predictors differed across clusters and between full-record and changepoint-year conditions: anthropogenic predictors were more influential in Clusters 1 and 2, whereas precipitation was more influential in Cluster 3. These results show that pumping nonstationarity across the HPAHOA is spatially heterogeneous, asynchronous, and regime-dependent.

Show More

 Contact

Resources
All 0
Collection 0
Resource 0
App Connector 0
Resource Resource
Asynchronous Regime Shifts in Groundwater Pumping in the High Plains Aquifer, 1960–2020
Created: May 12, 2026, 7:17 p.m.
Authors: Xu, Zherui · Yao Hu · Shihao Xi · Zhouyang Fu

ABSTRACT:

Groundwater pumping in agricultural aquifers reflects both hydroclimatic conditions and anthropogenic controls, but their relative importance can change over time. We analyzed annual pumping in the High Plains Aquifer Hydrologic Observatory Area (HPAHOA) from 1960 to 2020 using Deep Temporal Clustering, Bayesian changepoint detection, regime-aware XGBoost, and SHAP attribution. Approximately 13,000 cell-level pumping trajectories separated into three behavioral clusters, with asynchronous changepoints in 1974, 2000, and 2005. Models trained only on pre-changepoint data transferred poorly to post-changepoint periods, while models including post-changepoint information showed the greatest improvement in Cluster 3, indicating regime-dependent predictor–response relationships. SHAP attribution showed that dominant predictors differed across clusters and between full-record and changepoint-year conditions: anthropogenic predictors were more influential in Clusters 1 and 2, whereas precipitation was more influential in Cluster 3. These results show that pumping nonstationarity across the HPAHOA is spatially heterogeneous, asynchronous, and regime-dependent.

Show More