Hydrological modelling in snow- and glacier-fed catchments of the Upper Indus Basin by conceptual and machine learning approaches

TY - JOUR

T1 - Hydrological modelling in snow- and glacier-fed catchments of the Upper Indus Basin by conceptual and machine learning approaches

AU - Hassan Ougahi, Jamal

AU - Rowan, John S.

N1 - © 2026 The Authors.

PY - 2026/7/31

Y1 - 2026/7/31

N2 - Study region Upper Indus Basin (UIB), Himalaya and Hindu Kush. Focus of study This study evaluates the performance of a conceptually based glacio-hydrological model (GSM-HBV) and four data-driven models (DDMs), namely LSTM, RF, XGB and SVM, to simulate annualised, seasonal and flood flows for gauged catchments in the globally significant data-scarce catchments of the UIB. New hydrological insights or the region Five modelling schemes were evaluated across 23 snow- and glacier-fed catchments with the best-performing model identified based on achieving the highest ranking in at least two of the three evaluation metrics. In glacier melt catchments, XGB performed best for annualised flow (e.g., achieved the highest ranking in 4 out of 10 catchments) and GSM-HBV for flood events (e.g., 5 out of 10 catchments), whereas SVM, LSTM and RF were superior for predicting spring, summer and winter flows in 5, 4 and 6 out of 10 catchments respectively. In snowmelt dominated catchments, LSTM (e.g., 5 out of 13 catchments) was best for annualised flows, whilst SVM edged ahead of others for flood event prediction. Seasonally LSTM worked best for spring flows, whilst XGB, RF and GSM-HBV tied equally for summer and winter flows. The results show that no single model offers universally superior performance; thus, an ensemble approach is advanced whereby multiple solutions together provide robust ‘envelop’ predictions capturing over 90% of the observed flows.

AB - Study region Upper Indus Basin (UIB), Himalaya and Hindu Kush. Focus of study This study evaluates the performance of a conceptually based glacio-hydrological model (GSM-HBV) and four data-driven models (DDMs), namely LSTM, RF, XGB and SVM, to simulate annualised, seasonal and flood flows for gauged catchments in the globally significant data-scarce catchments of the UIB. New hydrological insights or the region Five modelling schemes were evaluated across 23 snow- and glacier-fed catchments with the best-performing model identified based on achieving the highest ranking in at least two of the three evaluation metrics. In glacier melt catchments, XGB performed best for annualised flow (e.g., achieved the highest ranking in 4 out of 10 catchments) and GSM-HBV for flood events (e.g., 5 out of 10 catchments), whereas SVM, LSTM and RF were superior for predicting spring, summer and winter flows in 5, 4 and 6 out of 10 catchments respectively. In snowmelt dominated catchments, LSTM (e.g., 5 out of 13 catchments) was best for annualised flows, whilst SVM edged ahead of others for flood event prediction. Seasonally LSTM worked best for spring flows, whilst XGB, RF and GSM-HBV tied equally for summer and winter flows. The results show that no single model offers universally superior performance; thus, an ensemble approach is advanced whereby multiple solutions together provide robust ‘envelop’ predictions capturing over 90% of the observed flows.

KW - Hydrological modelling

KW - Conceptual model

KW - Machine Learning models

KW - Snow- and glacier-fed catchments

KW - Flood prediction

KW - Upper Indus Basin

U2 - 10.1016/j.ejrh.2026.103796

DO - 10.1016/j.ejrh.2026.103796

M3 - Article

SN - 2214-5818

VL - 67

JO - Journal of Hydrology: Regional Studies

JF - Journal of Hydrology: Regional Studies

M1 - 103796

ER -