Closure to 'Estimation of Furrow Irrigation Sediment Loss Using an Artificial Neural Network' by Bradley A. King, David L. Bjorneberg, Thomas J. Trout, Luciano Mateos, Danielle F. Araujo, and Raimundo N. Costa

摘要

The authors wish to thank the discusser for his interest in the paper and taking the time to elucidate important elements of artificial neural network (ANN) modeling that, in retrospect, were overlooked by the writers in the original paper. Discusser Points 1 and 3 arise from division of the ANN data sets for training, testing, and validation in the paper. The importance of data division is often ignored or understated, with more attention given to model architecture and the learning algorithm (May et al. 2010). Data division should be conducted such that each data subset is representative of the others. Methods of data division include: (1) random division; (2) division such that statistical consistency of the subsets are retained; (3) division using self-organizing maps (SOM); and (4) division using fuzzy clustering (Shahin et al. 2004). Random division is easy to implement in software and the most common approach found in ANN software such as that used in the paper. The remaining methods likely need to be implemented external to the ANN software and enforced in the ANN software, if possible. The random division used in the paper did not insure statistical consistency among the data subsets, as is evident in the paper. However, statistical consistency among the data subsets can still result in significant variation in the ANN results depending upon which data are used for training, testing, and validation (Shahin et al. 2004).