Artificial neural networks (ANNs) adjust to the underlying behavior in the dataset using a training rule or optimizer. The most popular first-and second-order optimizers, Adam (AD) and Levenberg–Marquardt (LM), were compared with the aim of predicting extreme flash floods of a runoff-dominated hydrological system. A fully connected multilayer perceptron with a shallow structure was used to reduce complexity and limit overfitting. The inputs of the ANN were determined by rainfall–water level cross-correlation analysis. For each optimizer, the hyperparameters of the ANN were selected using a grid search and the cross-validation score on a novel criterion (PERS PEAK) mixing the persistency (PERS) and the quality of flood-peak restitution (PEAK). For an extreme and unseen event used as a test set, LM outperformed AD by 25% on all performance criteria. The peak water level of this event, 66% greater than that of the training set, was predicted by 92% after more training iterations were done by the LM optimizer. This shows that the ANN can predict beyond the ranges of the training set, given the right optimizer. Nevertheless, the LM training time was up to five times longer than that of AD during grid search.

We develop a holistic approach of watershed hydrology, focusing on the interfaces between critical zone components and relying on interdisciplinarity. We mainly explore the first Unsolved Problem in Hydrology (UPH): Is the hydrological cycle regionally accelerating/decelerating under climate and environmental change, and are there tipping points? and 5 other UPH (3,4,7,12,19). As feedbacks exist between the critical zone components, forming a complex system, it can lead to the emergence of tipping points and to shifts between hydrological regimes.A hydrological regime shift is indeed observed since 1950 in the Sahel (semi-arid region, Africa). Stream flows continuously increased in this region despite a severe regional drought in the 70s-90s. This suggests that a tipping point could have been surpassed, because of changes in climate and/or human practices. To explore this hypothesis, we develop a new modeling approach, relying on long term observations (1950-present) of hydrology, geomorphology and land used/land cover. We use System Dynamics modeling to capture feedbacks between water, soil structure, vegetation and flow connectivity between hillslope, channel and aquifer.The modeling results accurately represent the observed decade-scale evolution of the hydrological regime at several watershed scale (1-10000 km²). This shows that tipping points exist in semi-arid land hydrology and allows to explore which processes are at play. We identified domains within watersheds where regime shifts occurred at small scale, propagated across scales and led to large-scale shifts. Ultimately, we aim at identifying areas where the risk of an irreversible hydrological regime shift is high under various climatic and socio-economical possible futures.This study illustrates that using system dynamics to represent the critical zone is complementary to classical hydrologic modeling based on hydrodynamics. While the system-based approach represent correctly the decadal trends of the hydrological regime, the hydrodynamics approach did so for floods and drought at minute/year time scales.

The Sahel (the semi-arid fringe south of the Sahara) experienced a severe drought in the 70s-90s. During this drought, an hydrological regime shift was observed for most watersheds in the Central Sahel: runoff has significantly increased despite the rainfall deficit. Did the drought cause this regime shift ? What if the drought did not happen ? To answer these questions, we introduce a simple dynamical model that represents feedbacks between soil, vegetation and runoff at the watershed scale and at the annual time step. This model is forced with annual rainfall and evaluated using long-term observations of runoff from selected watersheds. We find that the model forced with observed rainfall reproduces well the observed regime shift in runoff. For the attribution of the regime shift to the drought, we rely on two sets of historical rainfall simulations from CMIP6 global climate models: fully-coupled simulations that do not reproduce the drought, and atmosphere-only simulations (AMIP) that represent the drought. Our results show that a regime shift would have been unlikely without the drought. This approach will be extended to identify areas that are likely to experience an hydrological regime shift in the future.