结合广义Gamma与卷积深度置信网络的PolSAR影像分类

PolSAR image classification based on the generalized Gamma distribution and convolutional deep belief network

针对极化合成孔径雷达(PolSAR)影像分类中标记样本较少导致分类模型难以拟合、未充分考虑地物统计特性等问题,该文提出了一种顾及广义伽马分布(GΓD)与空间邻域信息的极化SAR卷积深度置信网络(CDBN)的地物分类方法。通过引入深度置信网络与半监督学习方法,充分发掘已有数据信息,提高小样本条件下分类器的数据拟合能力,有效提高模型的分类精度。将GΓD的功率参数值域范围扩大到非零实数,增强GΓD对具有拖尾分布地物的描述;基于多极化信息,将GΓD与卷积结构有机融合到深度置信网络,其借助卷积结构的优点,更有效地描述不同类型地物的统计分布。基于EMISAR和GF-3全极化SAR影像进行分类实验,分别选取两个不同场景(农田和城市)作为实验区。结果表明,该方法在分类精度和统计分布拟合效果方面都优于传统方法。

In view of problems that the classification model is difficult to fit polarimetric synthetic aperture radar(PolSAR)images due to the small number of labeled samples and the statistical characteristics of ground objects are not fully considered,considering the generalized Gamma distribution(GΓD)and spatial information,a method for PolSAR image classification based on convolutional deep belief network(CDBN)was proposed in this paper.The existing data information were fully dug and the fitting ability of classifier was improved with small samples by the way of semi-supervised,and the classification accuracy was also improved effectively.The power parameter range of GΓD function was extended to non-zero real numbers to enhance the description of distribution features with tail.Based on multi-polarimetric information,the GΓD and convolutional network were integrated into the DBN to accurately describe the statistical distribution of different types of objects with the convolutional structure.Two different scenarios(farmland and city)were used as experimental areas based on the EMISAR and the GF-3 fully polarimetric SAR data.The results showed that the proposed method was superior to traditional methods in classification accuracy and statistical histogram fitting.