Abstract: In order to study the impact of multi-feature selection on remote sensing forest species classification with SVM, texture features at differrent scales of wavelet transform, four vegetation indexes and optimum band spectral features are selected to make up classification multi-feature vectors. Results show that the forest species classification accuracies with texture features, vegetation indexes and optimum band spectral features are the highest. They are respectively 84.4%, 86.5% and 91.0% for broadleaf, conifer and bamboo, 4.1%, 4.0%, 1.1% higher than those with only texture features, and 9.2%, 11.8%, 11.9% higher than those with only vegetation indexes. Generally speaking, the classification accuracies with multi-features are higher than those with single feature. Texture features in the multi-feature vectors could improve forest species separability obviously, and vegetation indexes have certain effectiveness. However, optimum band spectral features show weak effects on the raise of forest species classification accuracies.

Key words: forest species classification, remote sensing, Support Vector Machine（SVM）, multi-feature selection, wavelet transform

摘要： 为了研究遥感图像森林林型SVM分类多特征的选择对提高分类精度的影响，选取小波变换不同尺度纹理、四种植被指数、最优波段光谱特征等不同组合构成林型分类多特征向量进行分类。结果表明，纹理与植被指数、最优波段组合多特征的森林林型分类精度最高，阔叶林、针叶林和竹林的分类精度分别为84.4%、86.5%、91.0%，比纹理单类特征分类分别提高4.1%、4.0%、1.1%，比植被指数单类特征分类分别提高9.2%、11.8%、11.9%。多特征的分类精度一般要高于单类特征，纹理能够较明显提高林型可分性，植被指数也有一定的效果，但最优波段光谱特征的效果较弱。

关键词: 森林林型分类, 遥感, 支持向量机（SVM）, 多特征选择, 小波变换