Comparison between the effects of different training algorithms on the performance of MLP type statistical neural models for gas classification

Abstract

The present work uses a classification approach based on an artificial neural network (ANN) of the Multilayer Perceptron type (MLP). This algorithm was used to better discriminate individuals by highlighting non-linear relationships that are impossible to obtain with classical ordination methods. This method consists of projecting the spectrum of a gas, taken from remote sensing data, onto a three-dimensional space, using a MLP type neural network model. The latter adopts, during the training process, the gradient back-propagation algorithm, during which the mean squared error (MSE) at the output is continuously calculated and fed back to the input until it reaches a fixed minimum threshold, in order to correct the synaptic weights of the network. In this context, the ANN will provide undeniably effective solutions for classification. We have shown in this study that for the classification of gases (H2S-NO2 mixture, H2S, NO2), the best performing model is the one that uses as transfer functions, the Tansig function in the hidden layer and the Purelin function in the output layer, while using a Scalar Conjugate Gradient (SCG) training algorithm.