单变量特征选择的苏北地区主要农作物遥感识别

发布时间：2018-03-06 21:23

本文选题：单变量特征选择　切入点：光谱特征　出处：《遥感学报》2017年04期 　论文类型：期刊论文

【摘要】：遥感识别多源特征综合和特征优选是提高遥感影像分类精度的关键技术。农作物遥感识别中,识别特征的相对单一和数量过多均会导致作物识别精度不理想。随机森林(random forests)采用分类与回归树(CART)算法来生成分类树,结合了bagging和随机选择特征变量的优点,是一种有效的分类方法。单变量特征选择(univariate feature selection)能够对每一个待分类的特征进行测试,衡量该特征和响应变量之间的关系,根据得分舍弃不好的特征,优选得到的特征用于分类。本文基于随机森林和单变量特征选择,利用多时相光谱信息、植被指数信息、纹理信息及波段差值信息,设计多组分类实验方案,对江苏省泗洪县的高分一号(GF-1)和环境一号(HJ-1A)影像进行分类研究,旨在选择最佳的分类方案对实验区主要农作物进行识别和提取。实验结果表明:(1)多源信息综合的农作物分类精度明显高于单一的原始光谱特征分类,说明不同类型特征的引入能改善分类效果;(2)基于单变量特征选择算法的优选特征分类效果最佳,总体精度97.07%,Kappa系数0.96,表明了特征优选在降低维度的同时,也保证了较高的分类精度。随机森林和单变量特征选择结合的方法可以提高遥感影像的分类精度,为农作物的识别和提取研究提供了有效的方法。
[Abstract]:Multi-source feature synthesis and feature optimization of remote sensing recognition are the key technologies to improve the classification accuracy of remote sensing image. The relative singularity and excessive number of recognition features will lead to unsatisfactory crop recognition accuracy. Random forest random forestsuses the classification and regression tree cart algorithm to generate the classification tree, which combines the advantages of bagging and random selection of feature variables. Univariate feature selection) can test each feature to be classified, measure the relationship between the feature and the response variable, and discard the bad feature according to the score. The selected features are used for classification. Based on the random forest and single variable feature selection, this paper designs a multi-group classification experiment scheme based on multitemporal spectral information, vegetation index information, texture information and band difference information. The classification of Gaofen No. 1 (GF-1) and environmental No. 1 (HJ-1A) images of Sihong County, Jiangsu Province, were studied. In order to select the best classification scheme for the identification and extraction of the main crops in the experimental area, the experimental results show that the classification accuracy of the multi-source information synthesis is obviously higher than that of the single original spectral feature classification. It shows that the introduction of different types of features can improve the classification effect.) the optimal feature classification effect based on single variable feature selection algorithm is the best, and the overall accuracy is 97.07 and Kappa coefficient 0.96, which indicates that feature selection can reduce the dimension at the same time. The combination of random forest and single variable feature selection can improve the classification accuracy of remote sensing images and provide an effective method for crop identification and extraction.
【作者单位】：中国科学院遥感与数字地球研究所再生资源实验室;中国科学院大学资源与环境学院;
【基金】：国家自然科学基金(编号:41571422,41301497)~~
【分类号】：TP751

【参考文献】