基于RBF-SVM的草地高光谱图像分类  

Grassland hyperspectral image classification based on RBF-SVM

作　　者：赵烜赫 潘新[1] 马玉宝[2] 闫伟红[2] ZHAO Xuanhe;PAN Xin;MA Yubao;YAN Weihong(College of ComputeRand Information Engineering,InneRMongolia Agricultural University,Huhhot 010018,China;Institute of Grassland Research of CAAS,Huhhot 010020,China)

机构地区：[1]内蒙古农业大学计算机与信息工程学院,呼和浩特010018 [2]中国农业科学院草原研究所,呼和浩特010020

出　　处：《扬州大学学报（农业与生命科学版）》2022年第1期128-134,共7页Journal of Yangzhou University：Agricultural and Life Science Edition

基　　金：国家自然科学基金资助项目(61962048、61562067);中央级公益性科研院所基本科研业务费专项(1610332020020)。

摘　　要：高光谱图像具有光谱分辨率高、波段多、图谱合一等特点,可有效实现对草地快速无损分类,提高草地分类准确性。利用高光谱仪器(HyperSpec?PTU-D48E)采集可见-近红外光谱(400~1000nm)草地图像,采用多元散射校正(multiplicative scatteRcorrection,MSC)进行预处理;特征提取使用主成分分析(principal component analysis,PCA)白化法,选择最佳主成分作为支持向量机(support vectoRmachine,SVM)的输入,结合K折交叉验证法自动进行参数调优。比较不同SVM核函数对应的草地高光谱图像自动分类的识别结果,其中以基于高斯径向基核函数的支持向量机(RBF-SVM)分类结果较优,全局分类准确率(overall accuracy,OA)为98.89%,Kappa系数为0.99,分类时间为0.053098s,且优于梯度迭代决策树(gradient boosting decision tree,GBDT)及K近邻算法(K nearest neighboRalgorithm,KNN)。结果表明高光谱成像结合MSC-PCA白化-SVM(RBF)算法建立的识别模型可高效快速、准确无损地实现草地分类。Hyperspectral images have the characteristics of high spectral resolution, multiple bands, and image-spectrum merging, which can effectively achieve fast and non-destructive classification of grassland and improve the recognition accuracy. This study used a hyperspectral instrument(HyperSpec?PTU-D48 E) to collect the grassland image of visible-near infrared spectrum(400-1 000 nm), and the multiplicative scatter correction(MSC) was adopted for preprocessing. Then, feature extraction used principal component analysis(PCA) whiten method that selected the best principal component as the input of support vector machine(SVM), and automatically optimized parameters in combination with the K-fold cross-validation method. Then the recognition results of automatic classification of hyperspectral images of grassland corresponding to different SVM kernel functions was compared. Among them, the support vector machine based on Gaussian radial basis kernel function(RBF-SVM) was better, and the values of OA, Kappa coefficient and test time were 98.89%, 0.99 and 0.053 098 s, respectively. The RBF-SVM was also better than gradient boosting decision tree(GBDT) and K nearest neighbor algorithm(KNN). Experimental results show that the recognition model established by hyperspectral imaging combined with MSC-PCA whiten-SVM(RBF) algorithm can achieve grassland classification efficiently, quickly, accurately and non-destructively.

关 键 词：高光谱图像 草地分类 多元散射校正 主成分分析白化 支持向量机 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]