Daily Streamflow Simulations Improvement in Data Scarce Watersheds using different Optimization Techniques and Calibration Methods

Hydrological modeling is critical for effective water resources management, especially in developing countries such as Morocco where data are scarce. This study aims to improve daily river discharge predictions in 26 Moroccan catchments from 1993 to 2019. It evaluates the GR4J and MISDc models, focusing on optimizing their performances using four optimization techniques: Particle Swarm Optimization (PSO), the Nelder-Mead simplex algorithm (FMIN), Simulated Annealing (SA), and the Genetic Algorithm (GA). The two hydrological models are coupled with six calibration methods to provide the different ranges of uncertainties and to assess their consistency across diverse datasets. The methods include the split-sample or half-half method, the odd/even year method, as well as the calibration on a longer period than validation and vice versa. In addition, the Kling-Gupta Efficiency (KGE) and the relative bias were used as performance criterions. Due to the high elevation of some catchments studied and to the important amount of the snowmelt contribution in the river discharge at their outlets, a snow module incorporation was necessary to assess whether snowmelt impacts runoff or not. The outcomes demonstrate that all algorithms were able to successfully calibrate the GR4J and MISDc models (-0.26<median KGE< 0.34). However, FMIN and PSO demonstrated greater consistency in their performance across all calibration methods and proved to be the most computationally efficient algorithms, making them the best choices in situations requiring both time effectiveness and performance. Despite its slower speed, GA's robustness makes it a viable option under less time-sensitive conditions. The relative bias metric indicates that for the GR4J model, the FMIN, PSO, and GA had comparable and balanced performance, while SA showed greater variability. For the MISDc model, FMIN showed a tendency to slightly underestimate the discharge, while GA and PSO showed higher biases in some cases. In addition, MISDc significantly outperformed GR4J in simulating runoff across all catchments, making it a suitable choice for our region. The integration of a snow module in both models enhanced their performance in some larger pluvio-nival catchments, illustrating the complexity of snow dynamics in hydrological modeling and the need for high resolution data as well as ground measurements. Keywords: River discharge prediction, GR4J, MISDc, Moroccan catchments, Optimization methods, Data scarcity.