抄録
The multi-layer perceptron type artificial neural network model (MLP model) has been applied as a classifier in various classification fields, including remote sensing application. The back-propagation learning algorithm is widely employed in the MLP model as a simple learning algorithm. This algorithm, however, may lead to an unstable state of learning process, or take a lot of learning time in the learning process, if the learning rate is selected improperly. In this paper, a new learning rate institution equation is proposed for selecting of proper learning rate when the MLP model is employed in the classification of remote sensing satellite image data. At this time, a large scale training set is usually employed for learning of MLP model. Using the proposed institution equation, the learning rate is able to be directly determined by the training set size (number of training samples) and the network size (number of nodes) . At the same time, it guarantees that the learning process of MLP model is convergent. The proposed learning rate institution equation is more practical than conventional institution equations. The latter is often accompanied with a number of experiments, to obtain a proper learning rate. Using the proposed learning rate institution equation, not only a faster convergence is yielded but also a smaller error is obtained. To inspect the efficiency of proposed equation, two different size training sets are learned by two different MLP models respectively for the classification of multispectral image data.
