随机森林及其在遥感图像分类中的应用
发布时间：2019-07-09 18:07
【摘要】：随着遥感技术的不断发展，如何自动、准确、快速的为遥感图像分类一直是研究热点。由于红树林遥感图像训练样本获取困难，训练样本少，给自动化分类精度带来了很大的考验。本文在研究随机森林算法的基础上，结合TM图像的特征，提出了随机森林的改进算法，提升了自动化水平。 随机森林以其适用于小样本、稳定性强等特点被广泛应用于遥感分类，为了提升遥感分类的精度和效率，本文在随机森林作基础上作如下工作： 首先，为了提升TM遥感图像的分类精度，提出了完全随机的特征选取与组合的随机森林，能自动提取、挖掘TM图像中组合特征的信息。该算法是在特征线性组合的基础上加入了对特征组合个数的随机性和子特征空间大小的随机性，降低了随机森林的泛化误差，提升了分类精度。 其次，为了提升随机森林的分类效率，提出了基于克隆选择的随机森林，该算法引用人工免疫思想对随机森林进行选择优化，优化后，很好的压缩了随机森林，分类效率更高，分类精度也进一步提升。 最后，结合随机森林的性质，提出一个基于边缘最大化的未标签样本选取机制，实验证明，以该机制所选样本对提升随机森林的泛化能力有积极贡献。 为了验证算法的有效性，，所提算法都在UCI数据集上验证有效性，以保证通用性，并在遥感图像上和传统算法作进一步对比分析。
文内图片：
图片说明：TM图像图
[Abstract]:With the continuous development of remote sensing technology, how to classify remote sensing images automatically, accurately and quickly has been a hot research topic. Because of the difficulty of obtaining training samples of mangrove remote sensing images, there are few training samples, which brings a great test to the accuracy of automatic classification. Based on the research of random forest algorithm and the characteristics of TM image, an improved algorithm of random forest is proposed in this paper, which improves the automation level. Random forest is widely used in remote sensing classification because it is suitable for small samples and has strong stability. In order to improve the accuracy and efficiency of remote sensing classification, this paper does the following work on the basis of random forest. Firstly, in order to improve the classification accuracy of TM remote sensing image, a completely random feature selection and combination random forest is proposed, which can automatically extract and mine the information of combined features in TM images. On the basis of feature linear combination, the algorithm adds the randomness of the number of feature combinations and the randomness of the size of subspace, reduces the generalization error of random forest and improves the classification accuracy. Secondly, in order to improve the classification efficiency of random forest, a random forest based on clonal selection is proposed. The algorithm uses artificial immune idea to optimize the selection of random forest. After optimization, the random forest is compressed very well, the classification efficiency is higher, and the classification accuracy is further improved. Finally, combined with the properties of random forest, an untagged sample selection mechanism based on edge maximization is proposed. The experimental results show that the selected samples by this mechanism have a positive contribution to improving the generalization ability of random forest. In order to verify the effectiveness of the algorithm, the proposed algorithms are verified on the UCI dataset to ensure universality, and the remote sensing image is further compared with the traditional algorithm.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751
本文编号：2512337
文内图片：
图片说明：TM图像图
[Abstract]:With the continuous development of remote sensing technology, how to classify remote sensing images automatically, accurately and quickly has been a hot research topic. Because of the difficulty of obtaining training samples of mangrove remote sensing images, there are few training samples, which brings a great test to the accuracy of automatic classification. Based on the research of random forest algorithm and the characteristics of TM image, an improved algorithm of random forest is proposed in this paper, which improves the automation level. Random forest is widely used in remote sensing classification because it is suitable for small samples and has strong stability. In order to improve the accuracy and efficiency of remote sensing classification, this paper does the following work on the basis of random forest. Firstly, in order to improve the classification accuracy of TM remote sensing image, a completely random feature selection and combination random forest is proposed, which can automatically extract and mine the information of combined features in TM images. On the basis of feature linear combination, the algorithm adds the randomness of the number of feature combinations and the randomness of the size of subspace, reduces the generalization error of random forest and improves the classification accuracy. Secondly, in order to improve the classification efficiency of random forest, a random forest based on clonal selection is proposed. The algorithm uses artificial immune idea to optimize the selection of random forest. After optimization, the random forest is compressed very well, the classification efficiency is higher, and the classification accuracy is further improved. Finally, combined with the properties of random forest, an untagged sample selection mechanism based on edge maximization is proposed. The experimental results show that the selected samples by this mechanism have a positive contribution to improving the generalization ability of random forest. In order to verify the effectiveness of the algorithm, the proposed algorithms are verified on the UCI dataset to ensure universality, and the remote sensing image is further compared with the traditional algorithm.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751
【参考文献】
相关期刊论文 前3条
1 焦李成,杜海峰;人工免疫系统进展与展望[J];电子学报;2003年10期
2 王梦秋;万幼川;李刚;;核聚类改进的RBF神经网络遥感影像分类[J];测绘科学;2014年01期
3 王爱平;张功营;刘方;;EM算法研究与应用[J];计算机技术与发展;2009年09期
本文编号：2512337
本文链接：https://www.wllwen.com/guanlilunwen/gongchengguanli/2512337.html
