LSTSVM的样本缩减与空间信息融合方法

Sample reduction and spatial information fusion for LSTSVM

针对最小二乘孪生支持向量机(STSVM,least squares twin support vector machines)分类效率低的不足,在一对余(1-a-r)多分类器的基础上,提出一种基于样本缩减(SR)的LSTSVM(SRLSTSVM)分类算法。在核空间中通过距离计算,选出对分类超平面起决定作用的样本点,用于分类器的训练;与此同时,为了充分利用高光谱遥感图像的空间信息,通过主成分分析(PCA)和二维Gabor滤波获取像元的纹理特征,将高光谱遥感图像的空间信息和光谱信息在图像层进行融合用于分类。实验证明,本文提出的SR算法可以在不影响分类精度的基础上大大提高LSTSVM的分类效率,且结合空间信息后的LSTSVM的总体分类精度也有明显提高。

To overcome the low efficiency of least squares twin support vector machine （LSTSVM） in classifying,a new method called sample reduction LSTSVM （SR-LSTSVM） is proposed. The method greatly reduces the training samples in kernel space by calculating distance, but the ability of LSTSVM in classifying is not influenced. Meanwhile, to make full use of the spatial information, principal component analysis （PCA） and two-dimensional Gabor filter are adopted to get texture features of the pixel,and the hyperspectral images are classified by using the texture and spectral information. Experimental results show that the spatial-spectral information-based SR-LSTSVM classification algorithm can not only im- prove the classification effectiveness but also increase the overall classification accuracy.