Complementing data-driven and physically-based approaches for predictive morphologic modelling: results and implication from the Red River Basin, Vietnam

During the last 30 years, the delta of the Red River (Song Hong) in northern Vietnam experienced grave morphologic degradation. These processes can be attributed to a drastically altered sediment balance due to major upstream impoundments, sediment mining and land use changes, further aggravated by changing hydro-meteorological conditions. Resulting geomorphological alterations have severe effects on economic activities and endanger region-wide livelihoods. Rapidly progressing river bed incision, for example, threatens the irrigation of the delta’s paddy rice crops which constitute 20% of Vietnam’s annual rice production. Despite these severe impacts, river morphology was so far not included ¬¬¬¬¬¬into the current efforts to optimize basin wide water resource planning for a lack of suitable, not overly resource demanding modeling strategies. This paper assesses the suitability of data-driven modeling to provide insights into complex hydromorphologic processes and to complement and enrich physically-based modeling strategies. Hence, to identify key drivers of morphological change while evaluating impacts of future socio-economic, management and climate scenarios on river morphology and then resulting effects on key social needs (e.g. water supply, energy production and flood mitigation). Most relevant drivers and time-scales for the considered processes (e.g. incision) - from days to decades - were identified from hydrologic and sedimentologic time-series using a feature ranking algorithm based on random trees. The feature ranking pointed out bimodal response characteristics, with important contributions of long-to-medium (5 - 15 yrs.) and rather short (10d - 6 months) timescales. An artificial neural network (ANN), built from identified variables, subsequently quantified in detail how these temporal components control long term trends, inter-seasonal fluctuations and day to day variations in morphologic processes. Whereas the general trajectory of incision relates, for example, to the overall regional sediment balance over an extended time-horizon (>15 yrs.), upstream impoundments induce a much more rapid adaptation (1-5 yrs.). The applicability of the ANN as predictive model was evaluated by comparing its results with a traditional, 1D bed evolution model. The next decade’s morphologic evolution under an ensemble of scenarios, considering uncertainties in climatic change, socio-economic development and upstream reservoir release policies was derived from both models. The ANN greatly outperforms the 1D model in computational requirements and presents a powerful tool for assessing ensembles and quantifying uncertainties in river hydro-morphology. In contrast, the processes-based model provides detailed, spatio-temporally distributed outputs and validation of the ANN’s results for selected scenarios. We conclude that the application of both approaches constitutes a mutually enriching strategy for modern, quantitative catchment management. We conclude that the application of both approaches constitutes a mutually enriching strategy for modern, quantitative catchment management. We argue that physically based modeling can have specific spatial and temporal constrains (e.g. in terms of identifying key drivers and associated temporal and spatial domains) and that combining physically-based with data-driven approaches largely increases the potential for including hydro-morphology into basin-scale water resource management.