基于超像素与LightGBM的极化SAR图像地物分类

Polarimetric SAR image terrain classification based on superpixel and LightGBM

极化合成孔径雷达(SAR)图像的相干斑噪声降低了地物分类的准确率;联合极化SAR图像众多特征分类,过大的输入特征维度导致分类耗时长。为解决上述问题,提出一种基于超像素与LightGBM的分类算法。该算法充分利用极化SAR图像的极化特征与纹理特征,具备较强的分类能力;采用LightGBM处理大维度输入特征,能够快速得到基于像素的初级分类结果;利用SLIC生成基于超像素的极化SAR图像,并在各超像素内逐像素投票得到基于超像素的二级分类结果,抑制了相干斑的影响。利用实测极化SAR数据进行实验,分类的总体准确率超过97%,且分类耗时短。

Speckle noise of polarimetric synthetic aperture radar(SAR)image reduces the accuracy of terrain classification.Combine multiple features of polarimetric SAR image to do classification,and the large dimension of input features consumes too much time.To handle the above problems,we propose a classification algorithm based on superpixel and LightGBM.With polarimetric features and texture features,the algorithm is good at classification.LightGBM is used to deal with large dimension of input features,which can obtain the pixel-based first-level classification result efficiently.SLIC is used to generate the superpixel-based polarimetric SAR image,and the superpixel-based two-level classification result is obtained by voting pixel by pixel in each superpixel,which solves the influence of speckle noise.Experimental results,using the measured polarimetric SAR data,show that the overall classification accuracy is more than 97%,and it has a low time-consuming.