基于多类支持向量机的遥感图像分类及其半监督式改进策略

基于神经网络的遥感图像分类取得了较好的效果，但存在固有的过学习、易陷入局部极小等缺点.支持向量机机器学习方法，根据结构风险最小化(SRM)原理，表现出很多优于其他传统方法的性能，本研究的基于多类支持向量机分类器的遥感图像分类取得了达95.4％的分类精度.但由于遥感图像分类类别多，所需训练样本较大，人工选择效率较低，为此提出以人工选择初始聚类质心、C均值模糊聚类算法自动标注训练样本的基于多类支持向量机的半监督式遥感图像分类方法，期望能在获得适用的分类精度的基础上有效提高分类效率.

Multi-class SVM Based Remote Sensing Image Classification and Their Semi-supervised Improvement Scheme

Neural net based remote sensing image classification has obtained good results. But neural net has inherent flaws such as overfitting and local minimums. Support vector machine (SVM), which is based on Structural Risk Minimization(SRM), has shown much better performance than most other existing machine learning methods. Using multi-class SVM classifier high class rate of 95.4% is obtained. But for the class number of remote sensing image is much great, manually obtaining of training samples is a much time-consuming work. So a multi-class SVM based semi-supervised approach is presented. It is choosed that the initial clustering centroids manually first, then label the samples as the training ones automatically with fuzzy clustering algorithm. It is believed that this method will upgrade the classification efficiency greatly with practicable class rate.