Tan Kun, Du Pei-jun
Department of Remote Sensing and Geographical Information Science, China University of Mining and Technology, Xuzhou 221008, China.
Guang Pu Xue Yu Guang Pu Fen Xi. 2008 Sep;28(9):2009-13.
Based on the radial basis function neural network (RBFNN) theory and the specialty of hyperspectral remote sensing data, the effective feature extraction model was designed, and those extracted features were connected to the input layer of RBFNN, finally the classifier based on radial basis function neural network was constructed. The hyperspectral image with 64 bands of OMIS II made by Chinese was experimented, and the case study area was zhongguancun in Beijing. Minimum noise fraction (MNF) was conducted, and the former 20 components were extracted for further processing. The original data (20 dimension) of extraction by MNF, the texture transformation data (20 dimension) extracted from the former 20 components after MNF, and the principal component analysis data (20 dimension) of extraction were combined to 60 dimension. For classification by RBFNN, the sizes of training samples were less than 6.13% of the whole image. That classifier has a simple structure and fast convergence capacity, and can be easily trained. The classification precision of radial basis function neural network classifier is up to 69.27% in contrast with the 51.20% of back propagation neural network (BPNN) and 40. 88% of traditional minimum distance classification (MDC), so RBFNN classifier performs better than the other three classifiers. It proves that RBFNN is of validity in hyperspectral remote sensing classification.
基于径向基函数神经网络（RBFNN）理论和高光谱遥感数据的特点，设计了有效的特征提取模型，并将提取的特征连接到RBFNN的输入层，最终构建了基于径向基函数神经网络的分类器。对中国产的具有64个波段的OMIS II高光谱图像进行了实验，案例研究区域为北京中关村。进行了最小噪声分离（MNF），提取前20个分量进行进一步处理。将MNF提取的原始数据（20维）、MNF后从前20个分量中提取的纹理变换数据（20维）和提取的主成分分析数据（20维）组合成60维。对于RBFNN分类，训练样本大小小于整个图像的6.13%。该分类器结构简单、收敛速度快，易于训练。径向基函数神经网络分类器的分类精度高达69.27%，相比之下，反向传播神经网络（BPNN）的精度为51.20%，传统最小距离分类（MDC）的精度为40.88%，因此RBFNN分类器的性能优于其他三种分类器。这证明了RBFNN在高光谱遥感分类中是有效的。
