基于多目标遗传随机森林特征选择的面向对象湿地分类

发布时间：2018-03-05 05:11

本文选题：湿地分类　切入点：多光谱遥感影像　出处：《农业机械学报》2017年01期 　论文类型：期刊论文

【摘要】：以多时相Landsat8影像和SRTM DEM为数据源,对南瓮河流域进行了面向对象湿地分类。为削弱高维特征集对分类精度的影响,提出一种多目标遗传随机森林组合式特征选择算法(MOGARF)进行特征集优化。利用Relief F算法对完整特征集进行特征初选,再以基于随机森林的封装式多目标遗传算法进一步提取优化特征集。将所得特征集结合随机森林分类法提取湿地信息。并将结果分别与基于完整特征集和仅采用Relief F算法及Boruta算法提取的优化特征集的3种随机森林分类结果对比。试验结果表明,采用MOGARF算法特征选择后,特征维度降低至原来的10%,且分类精度最高,总体精度为92.61%,比其他分类方案提高0.35%~1.94%,Kappa系数为0.907 5,袋外误差为7.77%,比其他分类方案降低0.91%~1.48%。利用MOGARF特征选择的随机森林分类法是湿地分类的有效方法。
[Abstract]:Using multitemporal Landsat8 images and SRTM DEM as data sources, an object-oriented wetland classification was carried out in the Nanweng River basin, in order to weaken the effect of high Vitert collection on classification accuracy. A multi-objective genetic stochastic forest combination feature selection algorithm (MOGARF) is proposed to optimize the feature set. The complete feature set is selected by Relief F algorithm. Then the optimized feature set is extracted by the encapsulated multi-objective genetic algorithm based on random forest. The wetland information is extracted by combining the obtained feature set with the random forest classification method, and the results are compared with that based on the complete feature set and only by using the method of random forest classification. Three kinds of stochastic forest classification results obtained by Relief F algorithm and Boruta algorithm are compared. The experimental results show that, After feature selection with MOGARF algorithm, the feature dimension is reduced to 10%, and the classification accuracy is the highest. The overall accuracy is 92.61, which is 0.35% higher than that of other classification schemes. The coefficient of Kappa is 0.907 5, and the error is 7.77, which is 0.91% lower than that of other classification schemes. The random forest classification based on MOGARF is an effective method for wetland classification.
【作者单位】：吉林大学地球探测科学与技术学院;大连海事大学航海学院;
【基金】：东北地区国土资源遥感综合调查项目(85015B01009)
【分类号】：P237;P941

【相似文献】