改进的随机森林及其在遥感图像中的应用  

Improved random forests and its application to classification of remote sensing image

作　　者：姚明煌 骆炎民[1] 

机构地区：[1]华侨大学计算机科学与技术学院,福建厦门361021

出　　处：《计算机工程与应用》2016年第4期168-173,共6页Computer Engineering and Applications

基　　金：福建省自然科学基金项目(No.2012J01273)

摘　　要：对于遥感图像训练样本获取难的问题,引入适用于小样本分类的随机森林算法。为了随机森林能在小样本情况下有更优的分类效果和更高的稳定性,在决策树基础上提出了一种更加随机的特征组合的方法,降低了决策树之间的相关性,从而降低了森林的泛化误差;引入人工免疫算法来对改进后的随机森林进行压缩优化,很好地权衡了森林规模和分类稳定性、精度的矛盾。通过UCI数据集的实验表明,改进的随机森林的有效性及其优化的模型的可行性,优化后森林的规模降低了,且有更高的分类精度。在遥感图像上与传统的方法进行了对比。It is difficult to get training samples of remote sensing,Random Forests Algorithm（RFG）is introduced in the classification of remote sensing image. In order to improve the accuracy and stability of RFG on the case of small training samples,Firstly,a high random features selection method is proposed to build decision trees which have lower correlation,consequently a lower generalization error of random forests; secondly,Artificial Immune System is used to optimize and compress Random Forests,which can resolve the conflict between the accuracy and scale of Random Forest. The experiments are performed on the UCI data sets,and the results show that the efficiency and feasibility of the improvement algorithm and the optimization processing. After optimization,the forests have lower scale but higher accuracy. Finally,the improved RFG is used to remote sensing image and the results are analyzed by comparing to traditional methods.

关 键 词：遥感图像 随机森林 决策树 相关性 人工免疫 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]