Abstract: Compared with the multispectral remote sensing，the hyperspectral remote sensing can provide data with higher spectral resolution，and so more accurate classification of land cover is usually achieved.However，when the number of the training sample is equal with or less than the data dimension，the covariance matrices are close to singular or badly scaled and thus the maximum likelihood classifier will be degraded.The re-estimation of covariance matrices is necessary.Most methods of covariance estimation only select any two weighted items among the common covariance matrix，sample covariance matrix and their corresponding transforms，with ignorance of more combined items and their effects.Particle Swarm Optimization（PSO） is introduced to estimate covariance matrix.It investigates all the items through optimizing the weighting parameters.The classification results of hyperspectral data demonstrate that the proposed method is effective.

Key words: hyperspectral data classification, limited training samples, covariance matrix estimation, Particle Swarm Optimization（PSO）

摘要： 与传统的多光谱遥感相比，高光谱遥感具有更高的光谱分辨率，能更好地进行地物分类识别。但是，当训练样本数与数据维数相当，或小于后者时，会导致协方差矩阵近似奇异或奇异，使得经典最大似然分类失效，需要对协方差矩阵进行修正。典型的协方差阵估计方法往往只选取总体协方差、类别协方差及其相应变形中的两种形式进行组合，未考虑多种形式共同对协方差阵估计的影响。提出将PSO算法应用到协方差阵估计中，考虑所有形式的共同作用，对组合参数进行优化。最后，通过高光谱数据的分类实验证明了方法的可行性和有效性。

关键词: 高光谱数据分类, 有限训练样本, 协方差矩阵估计, 粒子群优化算法（PSO）