An Assessment of Internal Neural Network Parameters Affecting Image Classification Accuracy

Neural networks are attractive intelligence techniques increasingly being used to classify remote sensor imagery. However, their performance is contingent upon a wide range of algorithm and non-algorithm factors. Despite significant progresses being made over the past two decades, there is no consistent guidance that has been established to automate the use of neural networks in remote sensing. The purpose of this study was to assess several internal parameter's affecting image classification accuracy by multi-layer-perceptron (MLP) neural networks. The MLP networks have been considered as the most popular neural network architecture. We carefully configured and trained a set of neural network models with different internal parameter settings. Then, we used these models to classify an Enhanced Thematic Mapper Plus (ETM+) image into several major land cover categories, and the accuracy of each classified map was assessed. The results reveal that number of hidden layers, activation function, and training rate can substantially affect the classification accuracy and that a neural network with appropriate internal parameters can lead to a significant classification accuracy improvement for urban land covers when comparing to the outcome by the Gaussian Maximum Likelihood (GML) classifier. These findings can help design efficient neural network models for improved performance.