基于光谱与纹理特征的南疆盆地果树树种遥感识别研究

摘要

基于不同分辨率(2,8,16 m)遥感数据(GF-1),分别运用最大似然法、马氏距离、神经网络、支持向量机4种分类方法,结合光谱与纹理特征对南疆盆地4种主栽果树(核桃、红枣、香梨和苹果)进行遥感识别.结果表明,不同分辨率下,基于光谱与纹理特征的分类精度均比单纯依靠光谱特征参与下的分类精度有较大提高,2,8,16 m分辨率下分类总精度最高达到76．91％,67．46％,61．23％,比相应单纯依靠光谱信息分类提高了7．20％,5．00％,6．41％,Kappa系数也相应提高了0．14,0．03,0．06.不同分辨率下(2,8,16 m)识别4种果树的最佳纹理窗口分别为23×23、21×21、13×13窗口；对应的纹理特征组合为：方差、相异性、相关性；方差、均匀性、对比度；方差、对比度、相异性；支持向量机是识别4种林果树种的最佳分类方法,2m分辨率的遥感影像为识别4种果树的最佳选择.%Based on GF-1 remote data in different resolutions (2,8,16 m)the maximum likelihood meth-od,mahalanobis distance,neural network,suport vector machines combining spectrum and texture features were used to confirm 4 types of main cultivar fruit trees (walnut,j uj ube,pear and apple)in Southern Xin-jiang basin.The result indicated that classification accuracy based on spectrum and texture features im-proved more greatly than that based on only spectral characteristics at different resolutions.In the condition of 2,8,16 m resolutions,the highest total accuracy of classification increased to 76.91%,67.46%,61.23%respectively,which was increased by 7.20%,5.00%,6.41%,compared with that based on spectral informa-tion,Kappa coefficient increased correspodingly by 0.14,0.03,0.06.In the condition of 2,8,16 m resolu-tions,the best texture windows distinguishing four kinds of fruit tree were 23×23,21×21,13×13 win-dows,respectively.And the corresponding optimal texture combinations were variance,dissimilarity,corre-lation,variance,homogeneity,contrast,variance,contrast,dissimilarity at different resolutions.Support vec-tor machine was the best classification method for recognizing four kinds of forest fruit tree.2 m resolution remote sensing images was the best data sources of distinguishing four kinds of fruit tree.