Modeling rainfall-runoff process using artificial neural network with emphasis on parameter sensitivity

The gradient descent (GD) and Levenberg-Marquardt (LM) algorithms are commonly adopted methods for training artificial neural network (ANN) models for modeling various earth system and environmental processes. The performance of these algorithms is sensitive to the initialization of their parameters. The initialization of the algorithm's parameters for modeling different physical processes also varies process to process. However, there is a minority that tried to verify the sensitivity of the parameters of the algorithm than the sensitivity of the input data to the model. This work investigates the sensitivity of the popular ANN training algorithms to initial weights while modeling one of the earth system processes, i.e., the rainfall-runoff (RR) process. A novel methodology consisting of basic statistics for assessment of sensitivity of ANN parameters is proposed. The rainfall and flow data derived from three contrasting catchments are employed to establish the conclusions of this study. The results indicate that the RR model trained by LM algorithm is more robust in achieving performance with less variance irrespective of the existence of randomness in initialization of parameters than that of the GD trained models.