基于面向对象和随机森林模型的杭州湾滨海湿地植被信息提取

Classification of coastal wetland vegetation in Hangzhou Bay with an object-oriented, random forest model

湿地植被在湿地生态系统中起着无可替代的作用,其空间分布在很大程度上反映了滨海湿地的开发利用、生态环境特征和健康状况。以杭州湾南岸为研究区,以QuickBird影像和野外调查数据为数据源,基于面向对象原理在确定最优分割尺度的基础上采用随机森林模型,对滨海土地利用分类,并精确提取湿地植被。结果表明:面向对象和随机森林相结合的方法可以有效提取杭州湾5种湿地植被类型和6种土地利用类型,分类总体精度达86.90%, Kappa系数达到0.85, 5类滨海湿地植被的用户精度均达到85%以上,更有海三棱藨草Scirpus mariqueter的用户精度达到100%,充分说明了基于面向对象分割和结合随机森林模型方法适用于滨海湿地植被信息的精确提取。

Wetland vegetation plays an irreplaceable role in wetland ecosystems,and its spatial distribution largely reflects the distribution features and health information of wetlands.To explore the applicability of vegetation extraction in a coastal wetland and to achieve a higher classification accuracy,data sources from QuickBird and field survey data for the south bank of Hangzhou Bay were used with object-oriented image analysis and random forest modeling being combined.First,multi-scale segmentation of QuickBird images was carried out and evaluated by a quantitative global optimal judgment algorithm.Then,based on the image layer object at the optimal segmentation scale,multi-feature variables were fused into a random forest model to extract coastal wetland vegetation.Finally,the distribution map of wetland vegetation was analyzed,and accuracy of the classification result was evaluated.Results indicated that based on the random forest classification after optimal scale selection,five vegetation types and six other land use types could be effectively extracted.The final classification map was achieved with an overall accuracy of 86.90%and a Kappa coefficient of 0.85.Three typical vegetation types and their mapping precision(Scirpus mandshurica-100%,Reeds-95%,and Spartina alterniflora-76.7%)showed especially favorable results.Classification results of this coastal wetland vegetation,based on high-resolution remote sensing data for object-oriented image analysis and the random forest model that met accuracy requirements,could complete the coastal wetland vegetation information extraction.[Ch,4 fig.3 tab.34 ref.]