A hybrid machine learning-based multi-DEM ensemble model of river cross-section extraction: Implications on streamflow routing

Accurate representation of river channel and floodplain geometries is of prime concern for streamflow routing and flood inundation modelling under inadequate surveyed river cross-section data-condition, in many rivers across the globe. Due to global coverage and easy accessibility, the public domain medium resolution (1-arc second) DEMs such as SRTM, ASTER, and ALOS-AW3D30, with a linear vertical bias-correction approach, are widely used for hydrodynamic modelling of upland river networks. In this context, this study, for the first time, developed a hybrid machine learning-based Multi-DEM Ensemble (MDE) approach to simulate bias-corrected cross-sections using open source DEMs-extracted river cross-sections. Additionally, the hydraulic suitability of the simulated cross-sections and associated uncertainties were analysed. For this, the routing of streamflow and water level was performed using MIKE11-HD for a low-gradient flood-prone deltaic river basin of Eastern India (similar to 9000 km(2)) consisting of 19 river distributary systems. A comprehensive evaluation of the developed hybrid machine learning-based MDE models revealed an excellent performance of the Random Forest-Particle Swarm Optimisation-MDE (RF-PSO-MDE) model (NSE: 0.89-0.95; RMSE: 1.75-2.23 m) and Deep Neural Network-Particle Swarm Optimisation-MDE (DNN-PSO-MDE) model (NSE: 0.89-0.93; RMSE: 2.09-2.27 m) followed by the Support Vector Machine-Particle Swarm Optimisation-MDE (SVM-PSO-MDE) model (NSE: 0.87-0.92 and RMSE: 2.22-2.50 m). The MIKE11-HD simulations using RF-PSO-MDE and DNN-PSO-MDE model-simulated cross-sections were in close agreement with the observed streamflow and water level hydrographs, together with excellent matching of the flood peak, time to flood peak, and streamflow regimes. Predictive uncertainties analysis by the quantile regression technique confirmed an acceptable range of the Mean Prediction Interval and Percentage Interval Coverage Probability for the MIKE11-HD-RFMDE and MIKE11-HD-DNNMDE model variants including the MIKE11-HD-S model. Conclusively, the developed MDE models have potential for generating river cross-sections under sparse bathymetric conditions for the accurate simulation of streamflow and water level estimates using public domain digital elevation models.