《基于随机森林的遥感影像云影信息检测方法研究——硕士论文》由会员分享，可在线阅读，更多相关《基于随机森林的遥感影像云影信息检测方法研究——硕士论文（57页珍藏版）》请在金锄头文库上搜索。
1、分类号密 级UDC编 号 硕 士 专 业 学 位 论 文基于随机森林的遥感影像云影信息检测方法研究研究生姓名： 学号：指导教师姓名、职称： 专业类别（领域）：轻工技术与工程 年 月基于随机森林的遥感影像云影信息检测方法研究摘 要随着我国国产遥感卫星陆续发射成功并投入运行，遥感影像在越来越多的行业领域得到应用，发挥着重要的作用。光学遥感卫星传感器在成像过程中极有可能受到云层的干扰，遥感影像上经常会成对出现由云及其投影带来的不规则高亮区域以及暗淡的阴影区域，使相应的拍摄地域目标信息丢失或受到干扰。这降低了目标检测、识别和地物分类的准确性，对遥感影像的判读解译和应用带来不利影响。随机森林是一种基于决
2、策树的监督学习算法，在泛化能力和分类精度上均优于单一决策树，并且具有良好的鲁棒性，因此随机森林算法已成功应用于遥感影像分类。本文研究基于随机森林的遥感影像自动云影信息检测方法，设计实现影像云及云影信息的自动检测方法，本文主要工作为：（1）分析遥感影像中云、云影、地物的光谱和纹理特征，在传统的光谱、纹理特征基础上，加入了基于Sobel算子的边缘特征，以描述影像中边缘的幅度特性，构建有效区分云、云影、地物的特征组合。（2）利用遥感影像降采样得到的遥感影像快视图，采集云影区域、云区域、地物影像分块处理得到训练样本，提取样本特征集，利用随机森林算法训练影像分类器。（3）使用随机森林影像分类器对待测遥感
3、影像进行分类，得到云、云影、地物分类结果，接着对分类的各区块进行形态学闭运算，处理不同类区域相互重叠的情况，得到最终的分类结果，确定云和云影区域位置；（4）构建遥感影像测试集，通过改变随机森林训练参数进行分类器的训练与分类实验，确定最优的随机森林参数；针对分类器将较亮地物识别为云，将纹理平滑地物识别为雾的情况，增加对云、雾区域的“二次检测”；针对分类器将较暗地物识别为云影的情况，增加云影空间方位判定，以修正检测结果，提高检测精度；并对本文方法和优化改进后的方法进行精度评价。 本文在国产卫星的遥感影像测试集上进行云影信息检测试验，试验结果表明，本文方法可以较为准确地检测到遥感影像中的云影信息，具
4、有精度较高、运行速度较快、不依赖辅助数据的优点。关键词：随机森林；云影检测；影像分类；特征提取 AbstractWith the successful launch of remote sensing satellites in our country, remote sensing images are becoming ever more and more important in more and more industries. However, optical satellites will inevitably be affected by clouds and fogs in t
5、he process of imaging, the remote sensing images often appear in pairs from the cloud and its projection to bring the irregular highlight area and the dim shadows, so that the target information of the corresponding geographical area is lost or disturbed. This situation directly affects the accuracy
6、 of target detection, object identification and terrain classification, hinders the interpretation and application of remote sensing images. Therefore, detection and removal of clouds and cloud shadow information in remote sensing images are an important means to improve the quality of remote sensin
7、g images and the efficiency of data utilization.Random Forest is an effective supervised learning algorithm with high classification accuracy and good robustness, which has stronger generalization ability and classification effect than single Decision Tree. Therefore, the Random Forest algorithm has
8、 been successfully applied to classification application of remote sensing images.This thesis researched the automatic cloud shadow information detection method based on Random Forest algorithm, which utilized the computer to realize the automated detection of cloud and cloud shadow. To achieve the
9、research goal, the following work have been done in this paper:(1) The spectral and texture features of cloud, cloud shadow, and fog in remote sensing images are analyzed. On the basis of traditional spectral and texture features, edge features based on Sobel operator are added to describe the ampli
10、tude characteristics of the edges in remote sensing images, and the feature combination of cloud, cloud shadow and terrain is effectively constructed.(2) Segmentation processing of the fast view images, which was generated by down-sampling processing of corresponding original remote sensing data, wa
11、s be taken to extract the features of cloud shadow area, cloud area and terrain area sub-images. Then, train the shadow-cloud-fog image classifier with the sample features, using Random Forest algorithm.(3) Shadow-cloud-fog image classifier was utilized to classify the target remote sensing images a
12、nd obtain classification of the cloud, cloud shadow, and terrain images. Then, morphological block operations are performed on each images block of the classified area, then, cloud and cloud shadow area location was determined by area overlap detection.(4) Test set of remote sensing images was estab
13、lished, and the training and classification of the classifier tests are carried out by setting different Random Forest parameters to find the optimal parameters. In the case where the classifier recognizes the brighter terrain as cloud or the smooth terrain as fog, the secondary detection of cloud a
14、nd fog areas was added; in the case where the classifier recognizes the darker terrain as cloud shadow, spatial direction judging method was taken to correct cloud shadow area, and improve the accuracy of detection.The performance assessment of the cloud shadow detection method and optimized one whi
15、ch proposed by this paper was carried out under domestic satellite remote sensing images set. The experimental results show that the method in this paper detects the cloud shadow information in remote sensing image fairly accurately, and has the advantages of high precision, fast arithmetic speed, a
16、nd independent of auxiliary data.Keywords: Random Forests; Cloud Shadow Detection; Image Classification; Feature ExtractionIII目录摘 要IABSTRACTII1 概述11.1 研究背景及意义11.2 国内外研究现状分析21.2.1遥感影像云影信息检测研究现状分析41.2.2 随机森林方法在遥感影像处理中的研究现状分析41.3 本文主要研究内容62云、云影、雾影像特征提取方法72.1 遥感影像实验数据集72.2 云、云影、雾影像特征分析及提取82.2.1 灰度、颜色特征92.
