提出了一种利用独立分量分析(ICA)与支撑向量机(SVM)算法进行高光谱遥感影像分类的新方法。 采用ICA算法对高光谱遥感影像(PHI传感器获取, 80波段)进行了特征提取, 并以提取出的影像数据(光谱维数为20)构建SVM分类器。 对SVM算法进行核函数删选与参数寻优后, 发现采用RBF核的SVM算法(C=103, γ=0.05)分类结果最佳, 分类精度与Kappa系数分别达94.5127%与0.935 1, 优于BP-神经网络(分类精度39.4758%, Kappa系数0.315 5)、 波谱角分类(分类精度80.282 6, Kappa系数0.770 9)、 最小距离分类(分类精度85.462 7%, Kappa系数0.827 7)以及最大似然分类(分类精度86.015 6%, Kappa系数0.835 1)4种方法。 针对分类结果常出现的“椒盐”现象, 利用形态学算子对SVM(RBF核)分类结果进行了类别集群处理, 将分类精度与Kappa系数分别提高至94.758 4%与0.938 0, 获得了更接近实况的分类图像。 结果表明： ICA结合SVM算法准确率高, 是高光谱遥感影像分类的优选方法, 且类别集群是优化影像分类的有效方法之一。

A novel method was developed to classify hyperspectral remote sensing image based on independent component analysis (ICA) and support vector machine (SVM) algorithms. The characteristic information of the hyperspectral remote sensing image captured by PHI (made in China, with 80 bands) was extracted by ICA algorithm, and SVM classifier was established with the extracted image data (20 spectral dimensions). After kernel function selecting and parameter optimizing, it was found that the SVM algorithm(RBF kernel function; parameter C=103, γ=0.05)with accuracy 94.512 7% and kappa coefficient 0.935 1 has the best classification result, better than the results of four kinds of conventional algorithms, including neural net classification (accuracy 39.475 8% and kappa coefficient 0.315 5), spectral angle mapper classification (accuracy 80.282 6% and kappa coefficient 0.770 9), minimum distance classification (accuracy 85.462 7% and kappa coefficient 0.827 7) and maximum likelihood classification (accuracy 86.015 6% and Kappa coefficient 0.835 1). In order to control the “pepper and salt” phenomenon which appeared in classification map frequently, the classification result of SVM (RBF kernel) was operated by the method of clump classes using the morphological operators, and that the classification map closer to actual situation was acquired, with the accuracy and kappa coefficient increasing to 94.758 4% and 0.938 0, respectively. The study indicated that the ICA combined with SVM was an preferred method for hyperspectral remote sensing image classification, and clump classes was a effective method to optimized the classification result.