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.

Flash floods in urban and karst environments present major modeling challenges due to their complex hydrodynamics, characterized by a rapid urban runoff response and a delayed slower karst groundwater response. This study explores the use of artificial neural networks ANN (multilayer perceptron in particular) to predict flash floods at the downstream of the Las River in Toulon (France). The Las River is fed in a larger proportion by the nearby karst springs and in a smaller proportion by the urban drainage network. In this study, we propose an ensemble modeling strategy to address the system’s double hydrological regime. The initial step was to identify rainfall events in the six-year hydrometeorological database and classify them according to their karst contributions. Two specialized models: urban runoff model (UM) and karst model (KM) were trained solely on each event type (urban runoff and karst events). These models were combined by two methods in an attempt model all events, disregarding of their event type: the first approach was to combine the outputs of the specialized models in an ANN called output combination model (OM), the second approach was to combine the structures of the specialized models and retraining the model parameters called structure combination model (SM). A third more “basic” approach, called bulk model (BM), was to optimize the ANN by selecting the inputs with the best performance improvements. As expected, the specialized models (UM and KM) performed the best on the cross-validation sets and on the test sets on their respective event types but failed to generalize across regimes. The OM was the most robust ensemble strategy across all event types with consistent accuracy on predicting both urban runoff and karst flood events. The BM was better on the karst events while having worst performance on karst events and the SM was the least accurate model. These findings confirm the added value of combining specialized ANNs to model complex hydrological systems. In addition, selecting the right inputs to the models has a bigger impact on the model’s performance than choosing its structure by changing its hyperparameters.

Streamflow observation data is vital for flood monitoring, agricultural, and settlement planning. However, such streamflow data are commonly plagued with missing observations due to various causes such as harsh environmental conditions and constrained operational resources. This problem is often more pervasive in under-resourced areas such as Sub-Saharan Africa. In this work, we reconstruct streamflow time series data through bias correction of the GEOGloWS ECMWF streamflow service (GESS) forecasts at ten river gauging stations in Benin Republic. We perform bias correction by fitting Quantile Mapping, Gaussian Process, and Elastic Net regression in a constrained training period. We show by simulating missingness in a testing period that GESS forecasts have a significant bias that results in low predictive skill over the ten Beninese stations. Our findings suggest that overall bias correction by Elastic Net and Gaussian Process regression achieves superior skill relative to traditional imputation by Random Forest, k-Nearest Neighbour, and GESS lookup. The findings of this work provide a basis for integrating global GESS streamflow data into operational early-warning decision-making systems (e.g., flood alert) in countries vulnerable to drought and flooding due to extreme weather events.