结合光谱和形态剖面特征自适应加权影像分类

Adaptively weighted image classification based on spectral and morphological attribute profile features

针对遥感影像分类中多特征利用的问题,提出一种综合利用光谱和形态剖面特征的分类方法。该方法首先将原始影像经形态属性滤波器滤波所得的形态剖面特征和影像光谱值一起组成特征向量;然后依据训练样本各特征分量的方差确定不同地物类别相应特征分量的初始权重,并通过训练样本的特征加权获得各类别的初始中心;随后,根据初始权重计算每个训练样本到各聚类中心的加权距离,将最小距离对应的类别作为训练样本的初始类别;再对每一类别,根据该类训练样本中那些被错分为其他类别样本的均值与该类初始中心的距离来调整该类初始权重;最后,根据调整后的权重,采用加权距离最小分类方法对整个影像进行分类。实验结果表明,形态剖面特征有效地弥补了光谱信息区分能力不足的缺点,调整后的权重强化了区分能力强的特征分量。

To utilize muhiple features for classification of remote images, a novel classification algorithm adaptively combing spectral fea- tures and morphological attribute profiles is proposed. First, the morphological features are extracted from original image bands using multiple morphological attribute filtering procedures and are stacked together with spectral bands to create a spectral-spatial joint feature space. Second, the feature dimensionality for a certain land class is linked with an initial weight, which is determined by corresponding standard deviation of the training sample of the land class. Thereafter, the initial clustering centers are also calculated based on weighted training samples. Third, the weighted distances between every training pixel and initial clustering centers are calculated, and the pixel label with the nearest distance is taken as the class type. Fourth, the initial weights for each class are updated respectively according to the distances between the mean of the error-classified samples and the initial clustering centers of the corresponding class. Finally, every pixel is classified into the class with the weighted nearest distance. Experiment results demonstrate that the morphological attribute profile features provide the complementary information for recognizing spectral-similar land classes, and the adjusted weight can strengthen the feature components with strong classification ability.