Abstract: Remote sensing texture can be used to correctly classify forest species whose spectrums are similar. However, texture results in increase of the feature vector dimension and computation. Hence, TM forest image is classified with Support Vector Machine（SVM） based on texture of gray level co-occurrence matrix after it is decreasing dimension with Independent Component Analysis（ICA）. Results show that forest species can be classified correctly with the algorithm as its classification accuracy is 85.4% and Kappa coefficient is 0.73, which are greater than those with SVM, BP neural network, maximum likelihood and minimum distance algorithms. The accuracies to classify broadleaf, conifer and bamboo are 78.2%, 80.1% and 84.3%. Errors mainly result from mixed forests and intersection between two forest types. The classification accuracies of broadleaf, conifer are lower than that of bamboo because of easy appearance of broadleaf and conifer mixed forests.

Key words: remote sensing image, forest classification, texture, Independent Component Analysis（ICA）, Support Vector Machine（SVM）

摘要： 遥感图像纹理特征是光谱相近林型准确分类的有效方法，然而其带来分类特征向量维数增加和计算量增大。因此，对南方山区林地TM图像进行独立成分分析ICA降维，通过计算灰度共生矩阵获取纹理特征，使用SVM分类，研究林地类型的快速分类方法。结果表明，ICA与SVM法利用遥感图像纹理特征可较准确地实现林地类型分类，分类总精度、Kappa系数分别为85.4%、0.73，均高于SVM法、BP神经网络法、最大似然法、最小距离法；其对阔叶林、针叶林、竹林的分类精度依次为78.2%、80.1%、84.3%，误识率主要是由于混交林而造成两类林地之间存在交集，易出现的针阔混交林使得阔叶林、针叶林的分类精度低于竹林。

关键词: 遥感图像, 林地类型分类, 纹理特征, 独立成分分析, 支持向量机