利用无人机提取样本点的多源遥感影像分类方法研究

Land cover classification of multi-spectral remotesensing images by using Unmanned Aerial Vehicle(UAV) to extract marked sample points

摘要:

无人机遥感信息技术广泛应用于地表类型分类及监测。本研究均匀提取分布于研究区内的982个分类样本点,利用包括Sentinel-1/2(S1/2)数据集、S2数据集计算获得的植被指数(Vegetation index,VI)数据集和数字高程模型(Digital Elevation Model,DEM)数据集的遥感影像资料,借助随机森林分类模型完成研究区地表类型分类。结果显示:4种数据集组合(S2数据集、S2&VI数据集、S2&VI&DEM数据集,以及4个S2波段b3&b2&b4&b6组成的数据集)的制图总体精度(Overall Accuracy,OA)和Kappa系数精度都在75%和65%以上。此外,在不考虑位于边缘的样本点(即混合像元)后,分类制图结果具有更好的鲁棒,3种分类精度最高的数据集(S2数据集、S2&VI数据集和S2&VI&DEM数据集)OA和Kappa系数精度都提高至85%和79%以上。特别是kappa系数的精度更优,提高了近15%。本文研究结果为喀斯特地区的地表类型制图提供了准确而有效的技术手段。

Abstract:

UAV remote sensing information technology has been widely used in surface classification and monitoring.In this study,982 classification sample points from images captured by unmanned aerial vehicles evenly distributed within the study area were extracted.The land cover was classified using the random forest classification model based on the Sentinel-1 (S1) and Sentinel-2 (S2) datasets,vegetation index (VI) dataset estimated from the S2 dataset,and digital elevation model (DEM) dataset.The overall accuracy (OA) of the mapping and Kappa coefficient accuracy (κ) based on the DEM dataset were 74.54% and 61.73%,respectively,while the relevant values for the other four dataset combinations were ＞ 75% and 65%,respectively.In addition,removing the edge sample points (i.e.,mixed pixels) significantly improved the OA and κ of the three datasets with the highest classification accuracy to ＞ 85% and 79%,respectively.This study showed the effective method to distinguish weathered vegetation and bare land by automatic sample point extraction from aerial images.