Flow intermittence prediction using a hybrid hydrological modelling approach: influence of observed intermittence data on the training of a random forest model

Abstract

Rivers are rich in biodiversity and act as ecological corridors for plant and animal species. With climate change and increasing anthropogenic water demand, more frequent and prolonged periods of drying in river systems are expected, endangering biodiversity and river ecosystems. However, understanding and predicting the hydrological mechanisms that control periodic drying and rewetting in rivers is challenging due to a lack of studies and hydrological observations, particularly in non-perennial rivers. Within the framework of the Horizon 2020 DRYvER (Drying River Networks and Climate Change) project, a hydrological modelling study of flow intermittence in rivers is being carried out in three European catchments (Spain, Finland, France) characterised by different climate, geology, and anthropogenic use. The objective of this study is to represent the spatio-temporal dynamics of flow intermittence at the reach level in mesoscale river networks (between 120 and 350 km2). The daily and spatially distributed flow condition (flowing or dry) is predicted using the J2000 distributed hydrological model coupled with a random forest classification model. Observed flow condition data from different sources (water level measurements, photo traps, citizen science applications) are used to build the predictive model. This study aims to evaluate the impact of the observed flow condition dataset (sample size, spatial and temporal representativity) on the performance of the predictive model. Results show that the hybrid modelling approach developed in this study allows the spatio-temporal patterns of drying to be accurately predicted in the three catchments, with a sensitivity criterion above 0.9 for the prediction of dry events in the Finnish and French case studies and 0.65 in the Spanish case study. This study shows the value of combining different data sources of observed flow condition to reduce the uncertainty in predicting flow intermittence.