基于高分辨遥感数据的胡杨与柽柳树冠提取

胡杨、柽柳是干旱荒漠区生境的指示种,其树冠提取是荒漠生境遥感定量监测的基础。以塔里木河下游胡杨、柽柳为研究对象,基于QuickBird数据,使用光谱单数据源SVM、光谱结合纹理SVM、面向对象分类和最大似然分类法提取树冠。结果表明:1光谱结合纹理SVM比光谱单源SVM分类精度高9.65%,冠幅估测精度高7.18%,表明高分辨影像上纹理是提高分类精度的重要因素;2面向对象分类法精度最高,分类总体精度86.47%,较光谱单源SVM提高15.67%,较光谱结合纹理SVM提高6.02%,较最大似然法提高22.58%,其冠幅估测精度达87.45%。它兼顾面向对象影像分割与支持向量机方法优点,有效利用分割对象光谱、纹理和空间等信息,较好地解决了其他方法"同物异谱、异物同谱"造成提取树冠破碎的问题,使树冠提取具有较好的稳定性和较高精度。

Extraction of Tree-crown of Populus Euphratica and Tamarix Ramosissima based on High Resolution Remote Sensing Data

P.euphratica and T.ramosissimaare are indicator species of ecological environment in arid desert area,the extraction of their tree\|crown is the basis of quantitative monitoring of desert habitat by means of remote sensing.In this paper,Taking P.euphratica and T.ramosissimain in the lower reaches of Tarim River as the study object,the method of single data source SVM (Support Vector Machine)based on spectrum characteristics,SVM method based spectrum and texture characteristics,object\|oriented classification method and maximum likelihood classification method was used to extract tree\|crown from the QuickBird image.Single data source SVM method and maximum likelihood classification method is applied to classify the image which contains only spectrum characteristics.Other methods were carried out as follows:Firstly,calculating the textural measures by grey level co\|occurrence matrix and determining the optimal parameters for texture information by principal component analysis.Secondly,the optimal texture bands and the spectrum bands were combined into a new image.Finally,the support vector machine method and object\|oriented classification method was applied to classify the new image.The results show that:(1)The classification accuracy of SVM method based spectrum and texture characteristics is 9.65%,wich higher than that of single data source SVM method,the estimated accuracy of average crown diameter of the former is 7.18%,which higher than the later.The result indicates that texture is an important factor to improve the classification accuracy in high resolution images;(2)The tree\|crown extraction accuracy of object\|oriented classification is the highest.Its classification overall accuracy is 86.47%.The accuracy is 15.67% higher than single data source SVM method,6.02% higher than SVM method based on spectrum and texture characteristics,and 22.58% higher than maximum likelihood classification.Its estimated accuracy of average crown diameter is 87.45%,which suggests that object\|oriented classification method can effectively extract tree\|crown information in high\|resolution image and is better than the other classification methods.