A Knowledge-integrated Rbf Network for Remote Sensing Classification

Most Artificial neural networks (ANN) models used in the remote sensing classification are based on the multilayer perceptron (MLP) with back-propagation (BP) training algorithm. Compared to conventional statistical classifiers, MLP classifiers are non-parametric and distribution-free and is thus less restrictive in approximation, especially when distributions of features are strongly non-Gaussian. However, the major shortcomings of this class of networks are that they take relatively longer time to train and are prone to convergence to local minimum. The radial basis function (RBF) network, which combines the characteristics of the parametric statistical distribution model and non-parametric single layer perceptron, train much faster and are more stable than BP while keeping similarly complicated proximity. ANN, however, are very efficient in performing in classification tasks with low level of intelligence. They are less capable of reasoning with deep level knowledge, which is generally symbolic in nature. To better approximate the reality, integration of ANN with symbolic geographical knowledge is thus essential in the remote sensing classification. A knowledge-integrated RBF model that combines the power of approximation in high-dimensional space of the RBF network and the logic reasoning of rule-based inference is proposed in the present study. In addition to conceptual and technical discussions of the model, our arguments are substantiated by a real-life application. The experimental results show that the proposed model is more accurate, faster in training, simple in structure, and more interpretable.