Research on neural network model for remotely sensed image classification

In this Ph.D. dissertation, artificial neural network modd for remotely sensed image classification is studied. The \vork contains a survey of the existing learning algorithms related to the purpose of the work. Different learning control strategies to obtain appropriate solution on learning time and convergence improvements are proposed. The approach taken by the proposed method is to prevent the learning process from becoming instability.

Building upon the learning algorithm called self-built learning control mechanism focuses on the control of learning parameters automatically in backpropagation training is proposed. Interestingly, the proposed method require considerably less learning time and epoch than that required by the backpropagation algorithm. The influence of the automatic initial learning rate selecting on epoch is investigated. Further, the optimal value of learning control parameters is also experimentally determined.

A more specific learning control strategy, called a dynamic learning control algorithm is proposed. This algorithm investigates the effect of dynamic control strategies on the error surface conditio n. Moreover, to avoid the danger of learning rates growing very large, keeping the effective of learning rates in almost flat and steep regions. The addition of the minimum learning rate to prevent learning rate from decreasing to zero on the error surface has been considered.

The application of self-built learning control mechanism into a radial basis function network, called adaptive radial basis function algorithm, is proposed. The algorithm composes with a non-linear gaussian activation function in the hidden layer and the linear optimization of the output layer. It is found that the network constructed by the proposed method require considerably less learning time compared to the backpropagation and standard radial basis function.

A framework based on a hybrid neural network model using competitive learning algorithm is presented. This model is constructed by integration of the self-organizing map (SOM) and learning vector quantization (LVQ) method to enhance the quality of classifier decision regions. It is shown that, in comparison to the original LVQ method, the proposed hybrid model produced the best classification result. Abstract