用于高光谱遥感图象分类的一种高阶神经网络算法

High-Rank Artificial Neural Network Algorithm for Classification of Hyperspectral Image Data

BP神经网络近年来广泛地应用于遥感图象分类 ,但是它也有多层感知器神经网络的通病 ,即隐含层及其节点数问题、局部最小问题、训练速度问题等 .为了从根本上解决这些问题 ,该文提出了一种高阶神经网络分类算法 ,这种高阶神经网络没有隐含层 ,从而也就没有了隐含层及其节点数的问题 ;它的模式划分界面是非线性的 ,从根本上解决了局部最小问题 ;同时它的训练速度更快 ,分类精度更高 .该文详细介绍了这种高阶神经网络的构造、学习方法、模式分类方法和流程控制 ,并利用北京市沙河镇地区的高光谱数据进行了分类实验 ,取得了很好的结果 ,其训练样本和测试样本的分类精度都达到了 1 0 0 %,证明了这种高阶神经网络算法在结构 。

The BP neural network is widely used for classification of remote sensing image data nowadays. But it has the usual shortcomings of multilayer sensor neural network too: the question about the number of crytic layer and the number of crytic layer node, the question about local minimum, the question about training speed, and so on. In order to solve the questions thoroughly, a sort of classification algorithm of high rank neural network is developed in this research. This algorithm has not crytic layer, so it hasn't the question about the number of crytic layer and the number of crytic layer node. It's interface of model classification is nonlenear, so the question about local minimum is solved thoroughly. It's training speed is faster and the precision of model classification is greater than that of the BP neural network algorithm. In this article, the structure, flow chart and course control of this algorithm is introduced detailedly. Using the hyperspectral data in the destrict of Shahe town, Beijing city, an experiment is done and a excellent result is gained. The classification precision of training sample and the classification precision of test sample are all 100 percent. It is proved that the algorithm of high rank neural network has great advantages than other algorithms of neural network in structure, speed and precision.