Predicting performance of vegetative filter strips using artificial neural networks
The vegetative filter strips (VFS) are a best management practice. For quantifying the movement and amount of sediments and nutrients, the performance of VFS has to be modelled. Here artificial neural networks (ANN'S) were investigated, to model VFS and then compared to GRAPH. Data available from the literature and recent experiments were used. Artificial runoff was created. Flow samples were analysed for concentrations for total suspended solids, total phosphorus and soluble phosphorus, and particle size distribution. Available input-output data sets were used to train and test a multi-layered perceptron using back propagation (BP) algorithm and a radial basis function neural network using fuzzy c-means clustering algorithm. Sensitivity tests were done for finding optimum architectures of neural networks. The statistical analysis and comparisons between predicted and observed values for the three models showed that a BP network with 15 hidden units can model the performance of VFS efficiently, including the trapping of soluble P. They could predict the outputs, even without the particle size distribution. ANN'S have to be trained before being used to predict the outputs. GRAPH is mobile and could be successfully used for verification, since it takes into account the physical processes going on.