基于支持向量机的Landsat多光谱影像云检测算法研究

发布时间：2018-05-23 22:57

本文选题：Landsat卫星 + 多光谱图像　；参考：《安徽大学》2014年硕士论文

【摘要】：Landsat卫星影像广泛应用在资源调查、农业生产、环境监测、生态保护等领域。由于受到天气条件的影响,影像往往存在一些区域被云层覆盖,严重影响了影像的判读。准确地检测出Landsat卫星影像中的云层,对影像后续的分类、识别以及目标检测等进一步处理和应用具有重要的意义。本文通过提取Landsat卫星多光谱图像云层和地物特征,针对不同的云层检测应用需求,利用不同的支持向量机分类算法,对Landsat卫星图像的云检测问题进行了研究。主要的研究内容及研究成果包括以下几个方面： 1.介绍了Landsat卫星遥感图像云检测的研究目的与意义、国内外研究现状等。简介了支持向量机分类原理和孪生支持向量机算法。 2.针对单波段或者部分波段Landsat卫星多光谱图像,提出一种基于最小二乘孪生支持向量机的云检测方法。先根据云在不同波段中的大气辐射特点,结合Landsat7ETM+影像数据的光谱特性获得像元的光谱特征；再通过提取每个图像块的灰度共生矩阵得到相应像元点的纹理结构特征,以像元的光谱特性和纹理结构特征构造特征向量；最后利用最小二乘孪生支持向量机分类器进行Landsat7ETM+影像像元的云层检测,实验结果表明了方法的有效性。 3.针对多波段Landsat卫星多光谱图像,提出一种基于ACCA和WSVM相结合的云检测方法。首先利用ACCA方法对Landsat遥感图像进行云检测,将图像像元分成云像元、非云像元和待定像元；然后从Landsat遥感图像数据库中提取己知内容信息的图像像元点光谱特征作为WSVM的输入特征向量,通过建立关于训练样本中心距离的权重函数来获得样本的权重系数；最后利用改进的WSVM方法进行影像像元的云层检测。实验结果表明,该方法将待定像元中ACCA方法难以检测的半透明云检测出来。 4.研究了MLTK方法和半监督学习理论,并将其结合应用到遥感图像云检测中。在利用无标记签样本情况下,提出一种基于MLTK和STSVM的Landsat卫星遥感图像云检测方法,该方法先利用MLTK方法对Landsat图像进行云检测,将图像像元分成云像元和其他待定像元,再从Landsat遥感图像数据库中提取已知内容信息、未知内容信息的图像像元点光谱特征和纹理特征作为输入特征向量训练STSVM,构造最优分类超平面,最终将MLTK方法难以检测的薄云检测出来。仿真实验结果表明,与TSVM等方法相比较,该方法所获得的云检测结果在视觉效果上和定量评价上有明显提高。
[Abstract]:Landsat satellite images are widely used in resource survey, agricultural production, environmental monitoring, ecological protection and so on. Due to the influence of weather conditions, some areas are often covered by clouds, which seriously affect the interpretation of images. The accurate detection of clouds in Landsat satellite images is of great significance for the further processing and application of image classification, recognition and target detection. In this paper, the cloud and ground features of Landsat satellite multispectral images are extracted, and the cloud detection problem of Landsat satellite images is studied by using different support vector machine (SVM) classification algorithms for different cloud detection applications. The main research contents and results include the following aspects: 1. The purpose and significance of cloud detection in Landsat satellite remote sensing image are introduced. This paper introduces the classification principle of support vector machine and the algorithm of twin support vector machine. 2. A cloud detection method based on least square twin support vector machine (LS-TSVM) is proposed for multi-spectral images of single-band or part-band Landsat satellites. Firstly, according to the atmospheric radiation characteristics of cloud in different bands, combining with the spectral characteristics of Landsat7ETM image data, the spectral characteristics of pixels are obtained, and then the texture features of corresponding pixel points are obtained by extracting the gray level co-occurrence matrix of each image block. The spectral and texture characteristics of the pixel are used to construct the feature vector. Finally, the cloud detection of the pixel in Landsat7ETM image is performed by using the least square twin support vector machine classifier. The experimental results show the effectiveness of the method. 3. A cloud detection method based on ACCA and WSVM is proposed for multispectral images of multi-band Landsat satellite. Firstly, the image pixel is divided into cloud pixel, non-cloud pixel and undetermined pixel by using ACCA method for cloud detection of Landsat remote sensing image. Then the image pixel spectral feature of the known content information is extracted from the Landsat remote sensing image database as the input feature vector of WSVM, and the weight coefficient of the sample is obtained by establishing the weight function about the distance of the training sample center. Finally, the improved WSVM method is used for cloud detection of image pixels. The experimental results show that this method can detect the translucent cloud which is difficult to detect by ACCA method in undetermined pixels. 4. The MLTK method and semi-supervised learning theory are studied and applied to cloud detection of remote sensing images. In this paper, a cloud detection method for Landsat satellite remote sensing image based on MLTK and STSVM is proposed in the case of unlabeled sample. Firstly, cloud detection of Landsat image is carried out by using MLTK method, and the image pixel is divided into cloud pixel and other undetermined pixel. Then the known content information is extracted from the Landsat remote sensing image database. The pixel spectral feature and texture feature of unknown content information are trained as input feature vectors to construct the optimal classification hyperplane. Finally, the thin cloud which is difficult to detect by MLTK method is detected. The simulation results show that compared with the TSVM method, the cloud detection results obtained by this method are obviously improved in visual effect and quantitative evaluation.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】