利用高分二号数据提取香蕉林信息及精度分析

Extraction of Banana Orchards Based on GF-2 Satellite Imagery and Its Accuracy Analysis

针对海南农田地块细碎以及多云多雨气候条件下获取多时相的高质量卫星影像往往存在困难等问题,提出了一种利用单时相高分二号高分辨率卫星影像和随机森林算法的香蕉林信息提取方法。主要通过从高分辨率遥感影像中提取香蕉林的光谱和纹理等特征变量,然后利用综合不同光谱与纹理特征变量的随机森林分类算法进行香蕉林信息提取,并与以往的支持向量机分类算法进行了精度对比。结果表明,综合光谱和纹理信息的随机森林分类算法提取香蕉林空间分布结果最优,提取的香蕉林制图精度(PA)达到93.56%,用户精度(UA)达到87.43%;相比于支持向量机分类算法,PA和UA分别提高了11.99%和7.55%;相比只考虑光谱信息的随机森林分类算法,考虑纹理信息的随机森林分类算法提取的香蕉林PA提高了7.41%,UA提高了16.80%。研究结果可为人工园林的遥感信息提取提供技术参考。

It is usually difficult to extract land information of Hainan based on remote sensing data,because it is hard to obtain high quality remote sensing image under such rainy and cloudy weather,in addition the farmland in Hainan is finely divided.To solve these problems,in this paper,a method was proposed,which applied the algorithm of random forest and mono phase GF2high resolution remote sensing image to extract the information of banana orchards.Firstly,the spectral and texture feature variables of banana orchards were extracted from high resolution remote sensing image.Then,based on the spectral and texture information,use the random forest classification algorithm to extract banana orchards.Finally,the accuracy of extraction was compared with the result of support vector machine classification algorithm.The results showed that random forest classifier with multi feature information has the highest accuracy in the extraction of banana orchards,with the producer's accuracy(PA)was93.56%,and the user's accuracy(UA)was87.43%.Compared with SVM classifier,the PA and UA of banana orchards obtained by random forest classification were increased11.99%and7.55%respectively.Compared with spectral information,texture information were regarded as better factors to distinguish banana orchards and broad leaved evergreen forest,compared with the RF classification which only relies on GF2original spectral information,the PA and UA of banana orchards obtained from the random forest classification which integrated with texture information were increased7.41%and16.80%respectively.The results suggested that random forest classification algorithm,integrated with spectral and texture information,could effectively extract banana orchards,and the application of this method could provide technical support for extraction of other artificial garden spatial distribution.