Using a regionalized distributed hydrological modelling approach for prediction low flow on ungauged French territory based on the training of artificial neural network

This study delves into the exploration of low-flow rivers, a crucial subject within the global river network. These rivers, characterized by reduced flows over significant periods, play an essential role in various ecosystems. They constitute a substantial portion of the global river network, spanning diverse regions, including arid, semi-arid, temperate, humid tropical, boreal, and alpine areas. The flow variations observed in these watercourses are influenced by multiple factors, including climate change and increased water withdrawals associated with human activities. Ungauged basins, where reliable flow data is not readily available, present a significant hurdle in hydrological modelling. The absence of direct measurements makes it difficult to understand and predict the flow dynamics of rivers and streams, particularly in regions with low flow watercourses. To overcome this challenge, the study leverages the SMASH platform (Spatially-distributed Modelling and ASsimilation for Hydrology), a versatile multi-model framework capable of handling the complexities associated with ungauged territories.

The model implemented within the SMASH platform draws inspiration from the GR model family, a collection of global and semi-distributed models developed over the past years at INRAE. SMASH is a flexible, spatially distributed hydrological modelling platform capable of operating at high spatial and temporal resolution in both gauged and ungauged catchments. It is designed to simulate flow hydrographs across all grid cells in the computational domain.

Additionally, it incorporates functionalities for parameter sensitivity analysis and methods for both uniform and spatially distributed parameter calibration with different objective functions.

The principal aim of this study is to test the performance of various hydrological model structures, inspired by the GR model on the SMASH platform in low flow context. The evaluation centers on calibration and validation processes, employing uniform calibration techniques and regionalization approaches over a comprehensive dataset spanning 40 years at a daily time step. This extensive evaluation aims to elucidate the efficacy of these models in reproducing the low flows, seasonnality and bilan of watercourses over a set of basins (100) covering France with differents hydrometeorologic catchments. Furthermore, the study introduces a novel dimension by leveraging an artificial neural network (ANN) to process catchment descriptors specific to France. The ANN facilitates the exploration of regionalization by establishing a meaningful correspondence between select catchment descriptors and model parameters.

The study will then conclude with a comprehensive comparison of all simulations, highlighting the best hydrological model structure and regionalisation.