目的 深度置信网络能够从数据中自动学习、提取特征,在特征学习方面具有突出优势。极化SAR图像分类中存在海量特征利用率低、特征选取主观性强的问题。为了解决这一问题,提出一种基于深度置信网络的极化SAR图像分类方法。方法 首先进行海量分类特征提取,获得极化类、辐射类、空间类和子孔径类四类特征构成的特征集;然后在特征集基础上选取样本并构建特征矢量,用以输入到深度置信网络模型之中;最后利用深度置信网络的方法对海量分类特征进行逐层学习抽象,获得有效的分类特征进行分类。结果 采用AIRSAR数据进行实验,分类结果精度达到91.06%。通过与经典Wishart监督分类、逻辑回归分类方法对比,表现了深度置信网络方法在特征学习方面的突出优势,验证了方法的适用性。结论 针对极化SAR图像海量特征的选取与利用,提出了一种新的分类方法,为极化SAR图像分类提供了一种新思路,为深度置信网络获得更广泛地应用进行有益的探索和尝试。

Objective Several problems exist in polarimetric synthetic aperture radar (SAR) image classification, such as feature selection subjectivity and low utilization efficiency of massive features. Deep belief network (DBN) has a significant advantage in feature learning, which can be used in learning and extracting effective features from massive original features. Based on this observation, a polarimetric SAR image classification method based on DBN is proposed. Method The proposed method is capable of level-by-level learning and abstracting for the mass original polarimetric features. First, the original polarimetric feature sets are extracted from polarimetric SAR images. Second, 20000 samples are selected, and feature vectors are constructed. Each pixel contains 267 original polarimetric features and class labels. Thus, a pixel is a sample, namely, a feature vector. The feature vector is used as input in the DBN model. Then, the DBN model is built to extract abstract features, namely, effective features. These features are achieved through level-by-level learning. Finally, the logistic regression, a classifier at the top of the DBN model, is applied to classify the entire polarimetric SAR image. Result Considering AIRSAR data as an example, the overall classification accuracy can reach a high accuracy of 91.06%. The DBN method shows outstanding advantage in feature learning. Simulation experiments show that compared with the traditional Wishart supervised classification algorithm, the DBN algorithm performs much better in classification. Simultaneously, the necessity of the DBN model has been proven by comparing with the logistic regression classification. The logistic regression classification classifies the polarimetric image using the original polarimetric features without any deep learning and extraction. In brief, the effectiveness of the DBN model has been validated through analysis and comparison. Conclusion In this study, a novel polarimetric SAR image classification method is proposed. Mass polarimetric features of the polarimetric SAR image are utilized for the first time through the DBN. The advantages and applicability of the proposed method are analyzed. Overall, a novel method is proposed for polarimetric SAR image classification, which paves the way for further research and offer beneficial attempts for the utilization of DBN in polarimetric SAR image processing.