结合随机森林与K-means聚类评价湿地火烧严重程度  

Assessment of Wetland Fire Severity Using Random Forest Classifier and K-means Clustering Analysis

作　　者：林思美 黄华国[1] 陈玲[1] LIN Simei;HUANG Huaguo;CHEN Ling(Forest College,Beijing Forestry University,Beijing 100083,China)

机构地区：[1]北京林业大学林学院,北京100083

出　　处：《遥感信息》2019年第2期48-54,共7页Remote Sensing Information

基　　金：国家重点研发计划项目(2017YFC0504003-4);国家自然科学基金(41571332)

摘　　要：针对湿地植被存在典型的季节及年际变化特征,常用的遥感识别手段无法对湿地火烧严重程度实现准确评价的问题,提出了一种适用于湿地火烧严重程度的评价方法。基于2001年9月扎龙湿地的火灾事件,应用K-means聚类分析,从季节与年际2个方面的NBR阈值中获取不同火烧严重程度的训练样本,并利用随机森林机器学习算法建立基于光谱指数的分类模型,从而实现了湿地火烧严重程度的准确制图与评价。结果表明,交叉验证的分类总体精度为89.9%,各个火烧严重程度之间未出现严重的混分情况,且该模型具有一定的可移植性,能够成功地用于湿地火灾研究,从而为湿地火灾管理提供相应的参考依据。In view of the special seasonal and annual variation of wetland vegetation,the common remote sensing method is insufficient to assess wetland fire severity as accurate as possible.Therefore,an approach that is suitable for the assessment of wetland fire severity should be developed.This study presented a remote sensing approach to evaluate fire severity using the fire event information of Zhalong wetland in September 2001.The training samples of different fire severities were collected by K-means clustering analysis.The performance of nine spectral indices was assessed before implementation of the random forest machine learning algorithm.The 10-fold cross-validation result showed that the overall accuracy was 89.9%.There was no serious confusion among different fire severity levels,and the classification model can also be applied to another date with fire occurrences across the study area.Therefore,the remote sensing method that we presented can be successfully applied in wetland fire research,which could provide some valuable suggestions for wetland fire management.

关 键 词：湿地 火烧严重程度 随机森林 K-MEANS聚类 光谱指数 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]