基于聚类集成的半监督多/高光谱图像分类方法

Semi-supervised Classification of Multi/Hyperspectral Images Based on Cluster Ensemble

提出一种基于聚类集成的半监督多/高光谱图像分类方法。谱聚类是近年来出现的基于图论的、以相似性为基础的一类性能优越的聚类算法,能对任意形状分布的数据进行聚类,但对参数的变化比较敏感。聚类集成技术可有效提高单聚类算法的精度和稳定性,并具有良好的鲁棒性和泛化能力。算法利用聚类集成算法的优点并利用谱聚类的思想开发聚类集成算法的共识函数,将谱聚类作为聚类成员来构造聚类集成系统,使用高斯RBF核映射下的多维数据的光谱角制图计算权值矩阵W,并用Nystrm方法来降低算法的运算复杂度,实现了多/高光谱遥感数据的半监督分类。最后通过实验验证了该算法无论对多光谱还是高光谱都有较好的分类效果。

A semi-supervised, cluster-ensemble based method for the classification of multi / hyperspectral images is presented. Spectral clustering is a graph theory based clustering algorithm taking similarity as the basis, and has become increasingly popular in recent years. It can deal with arbitrary distribution of dataset but with a drawback for being sensitive to the scaling parameters. Cluster ensemble techniques are effective in improving both the robustness and the stability of the single clustering algorithm. Cluster ensemble also has a character of good robustness and generalization ability. The processing method in this paper utilizes the merits of cluster ensemble and develops a consensus function based spectral clustering algorthm. The clustering components are generated by spectral clustering. The affinity matrix is generated by computing the SAM between different datapoints. The Nystrm method is used to to speed up the classification process.Thus semi-supervised classification to multi / hyperspectral remote sensed data is completed. Experiments show that the method presented here has an excellent classificaation result for both multispectral and hyperspectral remote sensed dataset.