Neural networks for predicting nitrate-nitrogen in drainage water

Two artificial neural network (ANN) models, a trainable fast back-propagation (FBP) network and a selforganizing radial basis function (RBF) network, were developed for simulation of subsurface drain outflow and nitrate-nitrogen concentration in tile effluent. Experimental data collected at the Greenbelt Research Farm of Agriculture Canada over a 40-month period were used to train and validate the two models. The available field data were divided into training and testing scenarios, with the training file consisting of eight inputs and two outputs. A sensitivity analysis was performed by varying the network parameters to minimize the prediction error and determine the optimum network configuration. The best architecture for the FBP model comprised of 20 neurons in the hidden layer and a learning rate of 0.02, while the RBF network with a tolerance of 20 and a receptive field of 15 yielded 547 neurons in the hidden layer. Overall, the performance of the RBF neural network was superior to the FBP model in predicting the concentration of nitrate-nitrogen in drain outflow due to the application of manure and/or fertilizer. This information, in turn, can be used for proper fertilizer management; thereby, reducing not only the loss of valuable nitrogen fertilizer but also the potential for pollution of subsurface water by nitrate.