中文摘要：

高光谱遥感影像具有高维非线性、数据冗余多、训练样本难以获得等特点。在线性最小噪声分离变换MNF（Minimum Noise Fraction）的基础上，引入核方法，提出核最小噪声分离变换KMNF（Kernel Minimum NoiseFraction）高光谱遥感影像非线性特征提取方法。在KMNF特征提取后的影像上利用多类SVM进行高光谱影像分类，分析数据维数、样本个数对分类结果的影响，并与传统的最小距离分类方法进行对比。发现最小距离分类法存在维数灾难现象，当达到一定的特征维数之后，多类SVM分类方法受维数影响较小，具有一定的抗噪声能力，在一定程度上避免了维数灾难现象；利用多类SVM进行分类时，随着样本数目的减少，合理设置有关参数，高光谱图像的分类能够维持在较高精度；而传统的最小距离分类法当样本数量较小时，效果很差，这说明了SVM小样本分类的优势。

英文摘要：

Remote sensing images with hyper-spectrum have the characteristics of high dimensional nonlinearity, rich data redundancy, and difficult to obtain the training samples, etc. Based on linear minimum noise fraction （MNF） transformation, we introduce kernel method and put forward the nonlinear feature extraetion method with kernel minimum noise fraction （KMNF） for hyperspectral remote sensing image. Multi-class SVM is applied to the images with KMNF feature extraction performed for hyperspectral images classification, the impacts of data dimensionality and sample numbers on classification result are analysed and compared with the traditional minimum distance classification method. It is found that the dimension disaster phenomenon exists in minimum distance classification. After the feature dimension reaches a certain number, the multi-class SVM classification is less affected by the dimension and has the ability of noise-resistant, so it avoids the di- mension disaster phenomenon to certain extent. When using multi-class SVM in classification, with the reduction of sample number and by setting the parameters reasonably, the hyperspectral images can keep higher classification precision ; however for traditional minimum distance classification, when the sample number is small, the result would be very poor, this shows that SVM has the advantages in small-sample classification.