Impact of different saturation encoding modes on object classification using a BP wavelet neural network

Wavelet neural networks have been successfully applied to object classification due to their unique various advantages. The wavelet neural network used in this paper is a type of back-propagation algorithm-learning wavelet neural network. The log-sigmoid function and wavelet basis function satisfying the frame condition are employed as an activation function in the output and hidden layers, respectively, and the entropy error function is also used to accelerate the learning speed. The log-sigmoid function has two saturated values, 0 and 1, which are the value of the function at a point whose value changes slightly as the independent variable changes at a somewhat wide range. Using this property of the saturated values and simplifying the mathematical model of neural network classification, we may mathematically prove that using different saturated values to encode the modes can affect the training error, generalization ability, and anti-noise ability of the wavelet neural network, in turn resulting in differences in classification accuracy. The saturated and unsaturated value-encoding modes will both decrease the generalization ability of the network and reduce the classification accuracy due to excessively strong or weak anti-noise ability. Therefore, we propose a type of moderate saturated-value encoding mode, in which the anti-noise ability, the gradient, and error in training process are more moderate than the other two encodings, so that this kind of encoding mode can facilitate a stronger generalization ability and higher classification accuracy for the wavelet neural network, and which have been affirmed in the classification experiments of CHRIS remote-sensing imagery of the Huanghe estuary coastal wetland and SIR-C remote-sensing image of sea ice in the Labrador Gulf, and reaffirmed in classification experiments where noise was added to the test data.