Prediction of specific draft of different tillage implements:Using neural networks

Abstract

A Multilayer Perception with error backpropagalion learning algorithm was used to build neural network model to predict specific draft (kN/m) of different tillage implements from the field data. The neural network model was trained and tested with different sites, tillage implements, plowing depths, and forward operating speeds as input parameters and the measured specific draft as output parameter. The architecture of the neural networks consisted of two hidden layers with 24 nodes in the first hidden layer and 12 nodes in the second layer. The hidden and output layers have a sigmoid transfer functions in-neural networks model and the learning rule was momentum with 0.9 and step size 1.0. The best result was achieved at 65000 training runs that gave: minimum mean squared error equals to 0.0004 during training process. The results showed that the variation of measured and predicted specific draft was small and the correlation coefficient was 0.987 and mean squared error between measured and predicted specific draft was 0.1445.