【摘要】 红树林是最典型的滨海生态系统之一，红树林种间类型的精确识别对于红树林生态系统保护、修复及碳储量评估具有重要意义。遥感是开展红树林种间类型识别的有效手段，但传统的遥感红树林分类方法多是基于像元开展的，分类结果“椒盐”现象严重且精度还有很大提升空间。因此，本研究以东寨港红树林保护区为例，基于Sentinel-2 MSI影像，在传统遥感分类方法的基础上引入图像分割技术，分别构建了面向对象的支持向量机（Support Vector Machine,SVM）和随机森林（Random Forest,RF）分类法，并在此基础上对各模型的分类精度和适用性进行了分析。模型对比结果表明：（1）图像分割技术的引入能有效改善分类结果的“椒盐”现象，提升红树林种间类型的识别精度，基于像元使用SVM和RF分类算法总体分类精度分别可达78.82%(Kappa=0.75)和82.94%(Kappa=0.82)，面向对象的SVM和RF模型分类总体精度分别可达81.5%(Kappa=0.78)和92.67%(Kappa=0.88)，相较于以像元为分类对象的模型而言，后者精度分别提高了2.68%和7.43%;(2)从4个模型总体分类精度、各树种分类精度、模型稳定性和适用性方面来看，RF算法均优于SVM算法；（3）东寨港红树林分为6类，使用面向对象的随机森林分类，榄李和红海榄精度最高，其次为角果木，秋茄和无瓣海桑，海莲精度最低，为86.6%,6类树种分类精度均达85%以上。综上，基于面向对象使用随机森林分类算法构建分类模型可以准确识别分类红树林不同树种，为红树林种间精细化分类提供理论和技术支持。更多还原

【Abstract】 Mangroves are one of the most typical coastal ecosystems, and the accurate identification of mangrove interspecies types is of great significance for the conservation, restoration and carbon stock assessment of mangrove ecosystems. Remote sensing is an effective means to identify mangrove interspecies types. But traditional remote sensing methods for mangrove classification are mostly based on image elements, and there is much room for improving the accuracy of the classification results. Therefore, based on Sentinel-2 MSI images, this study introduced image segmentation techniques based on traditional remote sensing classification methods, and constructed the object-oriented Support Vector Machine(SVM) and Random Forest(RF) classification methods respectively. The analysis of the classification accuracy and applicability of four models demonstrates that:(1) The introduction of image segmentation technology can effectively improve the "salt and pepper" phenomenon of classification results and enhance the recognition accuracy of mangrove interspecies types, with the overall classification accuracy of SVM and RF classification algorithms being 78.82%(Kappa=0.75) and 82.94%(Kappa=0.82)respectively based on image elements, and the overall classification accuracy of object-oriented SVM and RF models being81.5%(Kappa=0.78) and 92.67%(Kappa=0.88), respectively. The overall classification accuracies of the object-oriented SVM and RF models were 81.5%(Kappa=0.78) and 92.67%(Kappa=0.88), respectively, with an improvement of 2.68% and 7.43%compared to the image element-based models;(2) the RF algorithm outperformed the SVM algorithm in terms of overall classification accuracy, classification accuracy of each tree species, model stability and applicability of the four models;(3)Dongzhaigang mangrove forest was classified into 6 categories. Using the object-oriented random forest classification, the highest precision was achieved for Lumnitzera racemosa and Rhizophora stylosa, followed by Ceriops tagal, Kandelia obovata and Sonneratia apetala. The lowest precision was 86.6% for Bruguiera sexangula, while the precision of all 6 categories reached over 85%. In summary, the object-oriented use of the random forest classification algorithm to construct a classification model can accurately identify and classify different tree species in mangroves, providing theoretical and technical support for interspecific refinement of mangrove classification.更多还原