文献类型：期刊文献

中文题名：荒漠绿洲区带状防护林遥感提取方法研究——以磴口为例

英文题名：Extraction Method of Oasis Shelterbelt Systems Based on Remote-Sensing Images——A Case Study of Dengkou County

作者：高峰[1,2,3] 姜群鸥[1,2,3] 辛智鸣[4] 肖辉杰[1,2] 律可心[1] 乔智[1]

第一作者：高峰

通信作者：Jiang, QO[1];Jiang, QO[2];Jiang, QO[3]

机构：[1]北京林业大学水土保持学院,北京100083;[2]北京林业大学水土保持与荒漠化防治教育部重点实验室,北京100083;[3]北京林业大学水土保持学院重庆缙云山三峡库区森林生态系统国家定位观测研究站,北京100083;[4]中国林业科学研究院沙漠林业实验中心,内蒙古磴口015200

年份：2022

卷号：42

期号：12

起止页码：3896-3905

中文期刊名：光谱学与光谱分析

外文期刊名：Spectroscopy and Spectral Analysis

收录：CSTPCD;;EI(收录号：20225213315649);Scopus;WOS:【SCI-EXPANDED(收录号:WOS:000910705800038)】；北大核心:【北大核心2020】；CSCD:【CSCD2021_2022】；PubMed;

基金：政府间国际科技创新合作重点专项(2019YFE0116500);中央高校基本科研业务费专项(2015ZCQSB03)资助。

语种：中文

中文关键词：防护林带识别;GF-2遥感影像;随机森林;面向对象

外文关键词：Shelterbelt extraction;GF-2 remote-sensing imagery;Random Forest;Object-oriented

分类号：S727.24

摘要：防护林是我国荒漠绿洲区主要植被类型,可为该地区防风固沙、水盐调控、水热平衡提供有力保障,调查防护林空间分布信息十分重要。然而荒漠绿洲防护林条带较窄、斑块面积小、分布广且零散,不易大尺度准确提取。为解决此难点,以磴口县荒漠绿洲为研究区,基于GF-2号遥感影像,采用面向对象分类技术提取防护林空间分布信息。分类前,首先基于局部方差(LV)和LV变化率(ROC)曲线,选取分割防护林的最优分割尺度。然后采用随机森林(RF)算法的袋外误差率(OOB error)及基尼系数(Gini)对包含光谱、形状和纹理的分类特征进行重要性评估并筛选特征、优化模型参数;最后,基于随机森林、CART决策树、支持向量机(SVM)、K近邻(KNN)四种分类器提取防护林空间分布信息并对比验证。结果表明:(1)采用ROC-LV曲线方法相比于遍历分割参数,可更客观更高效地筛选最优分割参数的可能值;(2)基于RF算法计算的袋外误分率和基尼系数可以有效筛除冗余特征,配合分类器参数优化,在保证分类精度的同时,有效提高分类器性能,大幅提升数据处理速度;(3)基于实测数据集对分类结果进行验证,结果显示基于随机森林算法的特征优化结合SVM分类器提取的防护林空间分布信息精度最高,生产者精度达到97.14%,总体防护林面积为208.99 km^(2),与实际210 km^(2)接近,在小区块中,SVM分类器的分类效果优于其他三种分类器;(4)因GF-2影像分辨率高,并且含有近红外波段,通过波段合成得到亚米级信息,故基于面向对象的方法能够以单条林带为基本单位研究防护林林网属性,例如提取断带信息等林网结构特征。研究结论可为荒漠绿洲区带状防护林提取提供重要技术支撑。
Shelterbelt systems are the main type of vegetation in the desert oasis regions,which provide a strong guarantee for wind-break and sand fixation,salt-water regulation and water-heat balance.It is important to investigate the spatial distribution information of shelterbelts.However,precisely mapping shelterbelts systems on a large scale are difficult due to narrow strips,small patches and wide&scattered distribution.This study aims to accurately map shelterbelts using object-oriented extraction based on GF-2 satellite imagery in Dengkou oasis.Firstly,the optimal scale parameter of SF segmentation was determined by local variance(LV)and rate of change(ROC)curve,and then the features space and classifier’s parameters were optimized by Out of bag error(OOB error)and Gini index through Random Forest(RF)algorithm prior to classification.Finally,Random Forest,CART decision tree,Support Vector Machine(SVM)and K-Nearest Neighbor(KNN)were compared and validated for shelterbelt systems extraction.The results showed that:(1)the ROC-LV curve method can obtain the possible value of optimal scale parameter more objective and more efficiently than iterating all scale parameter values.(2)OOB error and Gini index through RF algorithm can effectively eliminate the redundant features among spectral,shape and texture.The processing time was sharply reduced and ensuring the accuracy of the classification.(3)The classification results were verified based on the measured data sets,and the results showed that the feature optimization based on the RF algorithm combined with the SVM classifier was the best method for extracting the desert oasis shelterbelt systems,with the highest producer accuracy of 97.14%.Meanwhile,the extracted area of shelterbelt systems was 208.99 km^(2),which was close to reality(210 km^(2)).The SVM classifier performs better than the other three classifiers while zooming in a small areas;(4)Due to the high resolution of GF-2 images and the near-infrared band,sub-meter information can be obtained through appropriate band fusion.Based on the object-oriented method,a single shelterbelt can be used as the basic unit to explore the attributes and characteristics of the shelterbelts net.For example,the broken shelterbelts information could be extracted.All these conclusions will provideimportant technical support for the shelterbeltextraction in the desert oasis areas.