A simplified approach to quantifying predictive and parametric uncertainty in artificial neural network hydrologic models

[1] One of the principal sources of uncertainty in hydrological models is the absence of understanding of the complex physical processes of the hydrological cycle within the system. This leads to uncertainty in input selection and consequently its associated parameters, and hence evaluation of uncertainty in a model becomes important. While there has been considerable interest in developing methods for uncertainty analysis of artificial neural network (ANN) models, most of the methods are relatively complex and/ or require assumption about the prior distribution of the uncertain parameters. This paper presents an effective and simple way to perform uncertainty analysis for ANN-based hydrologic model. The method is based on the concept of bootstrap technique and is demonstrated through a case study of the Kolar River basin located in India. The method effectively quantifies uncertainty in the model output and the parameters arising from variation in input data used for calibration. In the current study, the uncertainty due to model architecture and the input vector are not directly considered; they have been minimized during the model calibration. The results from the case study suggest that the sampling variability of the training patterns as well as the initial guess of the parameters of ANN do not have significant impact on the model performance. However, despite good generalization properties for the models developed in this study, most of them fail to capture the hydrograph peak flow characteristics. The proposed method of uncertainty analysis is very efficient, can be easily applied to an ANN-based hydrologic model, and clearly illustrates the strong and weak points of the ANN model developed.