Improving daily streamflow forecasts in mountainous Upper Euphrates basin by multi-layer perceptron model with satellite snow products

摘要

This paper investigates the contribution of Moderate Resolution Imaging Spectroradiometer (MODIS) satellite Snow Cover Area (SCA) product and in-situ snow depth measurements to Artificial Neural Network model (ANN) based daily streamflow forecasting in a mountainous river basin. In order to represent non-linear structure of the snowmelt process, Multi-Layer Perceptron (MLP) Feed-Forward Backpropagation (FFBP) architecture is developed and applied in Upper Euphrates River Basin (10,275 km(2)) of Turkey where snowmelt constitutes approximately 2/3 of total annual volume of runoff during spring and early summer months. Snowmelt season is evaluated between March and July; 7 years (2002-2008) seasonal daily data are used during training while 3 years (2009-2011) seasonal daily data are split for forecasting. One of the fastest ANN training algorithms, the Levenberg-Marquardt, is used for optimization of the network weights and biases. The consistency of the network is checked with four performance criteria: coefficient of determination (R-2), Nash-Sutcliffe model efficiency (ME), root mean square error (RMSE) and mean absolute error (MAE). According to the results, SCA observations provide useful information for developing of a neural network model to predict snowmelt runoff, whereas snow depth data alone are not sufficient. The highest performance is experienced when total daily precipitation, average air temperature data are combined with satellite snow cover data. The data preprocessing technique of Discrete Wavelet Analysis (DWA) is coupled with MLP modeling to further improve the runoff peak estimates. As a result, Nash-Sutcliffe model efficiency is increased from 0.52 to 0.81 for training and from 0.51 to 0.75 for forecasting. Moreover, the results are compared with that of a conceptual model, Snowmelt Runoff Model (SRM), application using SCA as an input. The importance and the main contribution of this study is to use of satellite snow products and data preprocessing in ANN to improve the streamflow forecasts for ungauged or data sparse mountainous basins. (C) 2016 Elsevier B.V. All rights reserved.