Artificial Neural Networks Modeling of a Karstic Watershed in Mount Lebanon

Abstract

When applied to hydrology, the artificial neural networks (ANN) offer multiple advantages over conventional rainfall–runoff models. It was very interesting to assess the ANN performance over one of Mount Lebanon watersheds characterized by their nonlinear hydrologic regime due to the karstic nature, geomorphology and heterogeneous precipitation of rain and snow. The time delay neural network (TDNN) models in this study were assessed for their ability to simulate highly karstified watershed with little precipitation data, especially concerning snow contribution and for their ability to simulate fluctuated river flows. The selected watershed for this study was Nahr Ibrahim watershed with an area of 329 km² and an upper part located above 1700 m altitude taking part of the Cenomanian Plateau of Mount Lebanon. This karstic plateau is a stage for snow accumulation during winter and snowmelt during spring. The snowmelt is discharged by underneath karstic springs of Afqa and Roueiss, main contributors to Nahr Ibrahim flow. To achieve a better comprehension of the hydrologic regime, annual simulations with daily time step were conducted in this study. A simple snowmelt model was coupled with TDNN model (A) to makeup the lack of snow data. Model (A) which registered a considerable performance Nash criteria reaching 0.73 for the karstic springs and 0.66 for the watershed. However, two other methods were applied: the first, model (B), using BFI_max separation method which yielded a high performance of 0.95 and the second, model (C) with spring flows as input data which yielded 0.87. In this study, only nonlinear input output neural network was applied to avoid autoregressive models which would have definitely returned higher performances due to the long term rainfall–runoff correlation of the springs.