用于成像光谱数据特征矿物识别的人工神经网络结构研究

THE STRUCTURES OF ARTIFICIAL NEURAL NETWORKS USED FOR IMAGING SPECTRAL DATA PATTERN RECOGNITION

当某一问题很难甚至无法用数学方法建立精确模型时,人工神经网络的方法则显示了优势。对于一个具体问题,采用何种网络结构是至关重要的。本文以美国内华达州Cuprite矿区成像光谱数据特征矿物识别为例,采用6种不同结构的多层前馈网络模型,从其训练难度、运算效率及识别效果等方面进行了综合对比分析。

When it is difficult or even impossible to construct a precise model for solving a problem, the artificial neural networks (ANN) will show its advantage. The selection of the structure of ANN to deal with a specific problem is important. In this paper, six kinds of multilayer feedforward neural networks models were used for imaging spectral data pattern recognition of characteristic minerals, and their learning difficulties, operation efficiencies and recognition effects were studied synthetically.