遥感图象中波谱混迭象元分离方法的研究

提出用层次神经网络模型来解决遥感图象中波谱混迭象元的分离问题，即所谓的“同谱异构”问题，该模型由两级或多级神经网络级联而成，第一级神经网络主要用于波谱混迭象元的分类，采用第一个稳含层的BP网络，输入节点数目等于输入波段向量的维数，输出节咪数目等于期望类别数；第二级和后续层次的神经网络用于波谱混迭象元的分离，也采用只有一个稳含层的BP网络，其输入节点数目仍然等于波谱向量的维数，输出节点数目等于形成该混迭波谱的类别数，该模型可以高度精确地分离出波谱混迭的象元。

Investigation of a Methodfor Separating Spectral Overlap Pixels in Remote Sensing Images

Separating spectral overlap pixels in remote sensing images is a difficult problem known as "Different Land Covers with Same Spectrum". A hierarchical neural network model is presented to solve the problem, which consists of two or more levels of cascaded neural networks. The first-level neural network is mainly used for classifying spectral non-mixed pixels, using a BP neural network with one hidden layer. Its input node number is equal to the number of spectral bands and its output node number is equal to the number of the expected classes. The second or the higher levels of neural networks are used for separating the spectral overlap pixels, also using a BP neural network with one hidden layer. Its input node number is the same as the number in the first-level network and its output node number is equal to the number of the overlap classes. This model can separate spectral mixed pixels with very high accuracy.