Comparing conceptual and super ensemble deep learning models for streamflow simulation in data-scarce catchments

Sore and Masha river catchments in Baro Akobo river basin: Ethiopia. This research addresses the challenges associated with conventional data-driven streamflow modelling, which often exhibits inconsistent performance across different variability states. To bridge this gap, we explore the efficiency of model ensembles, a popular hydrological approach that harnesses the strengths of multiple models while preserving the fundamental characteristics of the data. Specifically, we compare three modified super ensemble meta-learners with eight single and hybrid machine learning base learners. The study also incorporates a semi-distributed HBV-light conceptual hydrological model and utilizes a decade of remote sensing and ground hydro-meteorological daily time series data. Different remote sensing-based vegetation index data products are employed to simulate single-step daily streamflow. The Recursive Feature Elimination (RFE) algorithm is applied to extract influential input parameters. Our findings consistently demonstrate that the three super ensemble learners outperform both the eight base models and the HBV-light model. The top-ranked Extra Tree Regression Super Ensemble (ETRSE) model exhibits superior performance, surpassing the HBV-light model by 24% according to the R2 performance measure. This study highlights the positive impact of selecting influential input parameters on the overall performance of machine learning models, providing valuable insights for hydrological modelling in the region.