Spatially transferable and interpretable machine learning for multi-horizon probabilistic early warning of hydrologic drought onset across U.S. watersheds
Year: 2026
Type: article
Abstract: Hydrologic drought early warning must identify imminent onset before low flow is already established, quantify probabilistic skill under severe class imbalance, and remain useful in hydroclimatically unfamiliar regions. We develop a daily multi-horizon framework using the U.S. Catchment Attributes and Meteorology for Large-sample Studies dataset (CAMELS-US), Daymet meteorology, and streamflow observations for 671 U.S. watersheds. Water years 1980–2004 serve as the training period, 2005–2009 as the validation period for model selection, and 2010–2014 as the held-forward test period. For onset, three model classes are compared using the same 53 engineered predictors and a common validation precision–recall area under the curve (PR-AUC) selection rule: ridge-penalized logistic regression, histogram gradient boosting (HGB), and a feed-forward multilayer perceptron (MLP). Held-forward receiver operating characteristic area under the curve (ROC-AUC) values for ridge/HGB/MLP are 0.841/0.883/0.897 at 7 days, 0.809/0.854/0.866 at 14 days, and 0.766/0.817/0.825 at 30 days; corresponding MLP PR-AUC-to-prevalence ratios are 5.15, 3.69, and 2.53. In five nested two-digit hydrologic unit code (HUC2) transfer folds, seven-day mean ROC-AUC is 0.825 for ridge, 0.866 for HGB, and 0.878 for MLP. Whole-basin paired bootstrap analysis confirms small but consistent MLP gains over HGB across onset horizons. Continuation shows a different ranking: pooled 30-day ROC-AUC is 0.792/0.832/0.828 for ridge/HGB/MLP, and pooled 60-day ROC-AUC is 0.803/0.848/0.832, with HGB also strongest at spell and basin-macro levels. Threshold-proximity and compact hydrologic benchmarks are weaker than all three onset models, while drought-definition audits show that strong onset discrimination persists under seasonal 10th- and 5th-percentile flow thresholds (Q10/Q5), fixed annual Q20/Q10/Q5 thresholds, and one- or two-day interruption pooling. Thresholded alert diagnostics nevertheless produce substantial false-alert burden, emphasizing that discrimination is not equivalent to a deployable warning policy. These results support spatially transferable probabilistic early warning and show that the value of added model flexibility depends on the forecasting task.
Authors: Maha Moussa, Yan Sun
Institutions: Utah State University, Benha University
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,090
OpenAlex ID: W7227371125
