无人机遥感影像面向对象分类的冻土热融滑塌边界提取

Object-oriented classification of unmanned aerial vehicle image for thermal erosion gully boundary extraction

全球气候变暖及人类活动导致青藏高原大面积冻土退化、热融滑塌等问题,严重影响了多年冻土区工程建设和生态环境。利用无人机高空间分辨率影像和面向对象分类技术进行了黑河上游俄博岭垭口冻土区热融滑塌监测实验,详细分析了最邻近、K-最邻近、决策树、支持向量机(support vector machine,SVM)和随机森林5种面向对象监督学习方法提取冻土热融滑塌边界的性能和精度,并使用野外实测数据对实验结果进行验证。结果表明,面向对象分析中分割尺度对热融滑塌提取结果影响较小,而不同组合的分类特征影响较大,因此选择合适的分类特征是关键;5种分类方法的总体精度均在90%以上,其中SVM方法的Kappa系数高于其他4种分类方法,表明该方法在本次实验研究中更适合无人机遥感影像冻土热融滑塌边界提取。无人机高空间分辨率遥感与面向对象分类方法相结合在冻土热融滑塌监测中具有广阔的应用前景。

Global climate warming and human activities have caused large areas of permafrost degradation and thermal erosion gully in the Tibetan Plateau,seriously affecting the engineering construction and the ecological environment in permafrost regions.In this study,high resolution unmanned aerial vehicle (UAV) images and object-oriented classification approaches were applied to extracting the thermal erosion gullies in Eboling Mountain of Heihe River.Five kinds of object-oriented supervised learning algorithms,namely nearest neighbor,K-nearest neighbor,decision tree,support vector machine (SVM),and random forest,were analyzed for the capability and accuracy of the extraction of thermal erosion gullies in detail.The field GPS data were used for evaluating the classification accuracy.The results show that,in the object-oriented image analysis,the segmentation scale parameters have little effect on the extraction of thermal erosion gullies,wheres classification features have a greater impact,so it is important to select the appropriate classification features.The overall accuracies of the five machine learning methods are all over 90%,among which the Kappa coefficient of the SVM is higher than the other four classification methods.This means that SVM is more suitable for the thermal erosion gullies boundary extraction of UAV images in this study.The combination of high resolution UAV images and object- oriented classification methods has broad application prospects in the extraction of the thermal erosion gullies.