将关联向量机应用于高光谱影像分类,实现高维空间中训练样本不足时分类器的精确建模。从稀疏贝叶斯理论出发,分析关联向量机原理,探讨一对多、一对一和两种直接的多分类方法。实验环节比较了各种多分类方法,并从精度、稀疏性两方面将关联向量机与支持向量机等经典算法比较。实验结果表明,两种直接的多分类方法内存占用大、效率低;一对多精度最高,但效率较低;一对一计算效率最高,精度与一对多近似。关联向量机精度不如支持向量机,但解更稀疏,测试样本较多时实时性好,适合处理大场景高光谱影像的分类问题。 The relevance vector machine is applied to the classification of hyperspectral images to realize the accurate modeling of classifiers in the case of insufficient training samples in high-dimensional space. Based on the theory of sparse Bayesian, the principle of relevance vector machine is analyzed and one-to-many, one-to-one and two direct multi-classification methods are discussed. In the experimental section, we compared various multi-classification methods and compared the classical algorithms such as relevance vector machine and support vector machine from two aspects of accuracy and sparseness. The experimental results show that the two direct multi-classification methods have large memory usage and low efficiency. The one-to-many precision is the highest, but the efficiency is low. The one-to-one calculation is the most efficient and the accuracy is close to one-to-many. The accuracy of SVM is lower than that of SVM, but the solution is more sparse and has better real-time performance when there are more test samples, which is suitable for the classification of hyperspectral images in large scenes.