Bias correction of Global Flood Awareness System (GloFAS) data for multi-station river flow prediction by ensemble modelling

Abstract

The Global Flood Awareness System (GloFAS) is a promising flood-modelling tool available for most river basins worldwide. However, it contains inherent biases, notably linked to its spatial resolution. This study proposes a multi-station ensemble modelling approach to regionalize and improve GloFAS-based river flow predictions. We used daily GloFAS-ERA5 data for the Sahzab, Mirkuh, and Markid flow stations in the Ajichai catchment, northwest Iran. A range of Machine Learning (ML) models were examined, including shallow learners, Feedforward Neural Network (FFNN), Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Regression (SVR), and the DL Long Short-Term Memory (LSTM) model. We tested these models and an ensemble of shallow learners under several scenarios that used both raw and bias-corrected GloFAS inputs. Precipitation and temperature from local weather stations and European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) were used as inputs. Observed discharge from 1993 until 2024 served as the target. In the multi-station strategy, upstream discharge from Sahzab and Mirkuh was included to improve downstream predictions at Markid. Integrating GloFAS data with advanced ML techniques, particularly ensemble learning within a multi-station framework, noticeably improved discharge prediction accuracy. Results, evaluated with Root Mean Square Error (RMSE) and the Coefficient of Determination (DC), show that the proposed Neural Averaging (NA) nonlinear ensemble outperforms individual ML models. At the Markid station, the multi-station approach improved performance over the single-station setup: RMSE decreased by ≈2.2% during calibration and ≈9.4% during verification, while DC increased by ≈3.9% and ≈25%, respectively. These improvements can support more reliable local flood forecasting and water-management decisions, and thus inform regional water policy and risk-reduction planning.