基于支持向量数据描述的遥感图像病害松树识别研究

发布时间：2018-05-04 15:21

本文选题：无人机 + 遥感图像　；参考：《安徽大学》2014年硕士论文

【摘要】：中国是世界上林业有害生物灾害最严重的国家,而松树又是我国非常重要的森林物种资源,具有很高的经济和实用价值。近年来,由于受松材线虫病的侵染,许多地区的松林相继发现大量的枯死松树。由于该病害具有传播速度快、致病能力强、适生范围广等特点,导致整个松林遭到毁灭,因而被列为林业病虫害之首,不仅严重威胁我国林业资源和生态建设,而且给国民经济带来了巨大的损失。因此,研究实现松林中病害松树的识别与监测,对预防松树病虫害、保护松林资源和促进经济社会可持续发展等具有十分重要的意义。传统的病害松树识别主要依靠病虫害专家人工实地调查进行,受山区地形条件、调查人员责任心等影响,这种方法很难及时准确地掌握疫情,而且难以大范围推广。随着遥感技术的快速发展,利用卫星监测平台获取高分辨率遥感数据对监测大面积枯死松树非常有效。然而受卫星运行周期、大气环境等因素的影响,卫星遥感技术难以发现单株染病松树,另外受经费成本的制约,不适合长期大范围地应用。因此寻找一种既经济又有效的识别与监测方式,已成为重要的研究课题。本文以小型无人机为空中飞行平台,以双光谱相机为遥感探测手段,获取松林的可见光和近红外图像,综合运用图像处理、模式识别等知识实现病害松树的识别与监测。主要研究内容如下： 1.阐述了松树病虫害识别与监测的目的和意义、国内外研究现状、研究区与数据源,并介绍了支持向量数据描述的基本理论和发展现状。 2.提出了一种基于加权支持向量数据描述的遥感图像病害松树识别方法。首先根据不同地物的特点,提取可见光图像和近红外图像各颜色分量作为相应像素点的颜色特征,接着通过提取加窗图像块的灰度共生矩阵得到中心像素点的纹理特征,然后通过建立关于训练样本中心距离的权重函数来构造加权支持向量数据描述多分类模型,进而将图像中地物多输出分类,从而实现病害松树的有效识别。多组对比实验验证了该方法的识别准确性。 3.提出了一种基于多特征和改进的加权支持向量数据描述多分类的病害松树识别方法。该方法基于每类样本的单像素和中心局部区域进行特征提取,构建相应特征向量,在传统支持向量数据描述多分类方法的基础上,将处于分类模型模糊区域的样本使用K近邻隶属度函数估计法进行判别决策,并结合权重函数和小波核函数的优点,构建加权小波支持向量数据描述多分类器,并应用于病害松树的识别。实验结果表明,该方法避免了传统支持向量数据描述多分类方法中对模糊区域样本决策的盲目性,在局部样本中寻求密度高的样本类别,大大提高了病害松树的识别率。 4.提出了一种基于特征稀疏表示和加权小波支持向量数据描述的松树多级病害识别方法。首先对松树双光谱图像进行特征提取,然后利用特征数据集构造学习字典,根据学习字典计算每个样本的稀疏系数,得到样本的稀疏表示特征向量,输入加权小波支持向量数据描述多分类器模型,从而将受不同病害程度的松树有效地分类识别。实验结果无论从视觉上还是定量评价上均说明了该方法的可行性和有效性。
[Abstract]:China is the most serious forest pest in the world, and the pine tree is one of the most important forest species resources in our country. It has high economic and practical value. In recent years, because of the infection of pine wood nematode disease, a large number of dead pine trees have been found in many areas. Because of the disease, the disease has the speed of rapid transmission and the pathogenic energy. The characteristics of strong force and wide range of life have caused the whole pine forest to be destroyed, so it is listed as the first of the forestry diseases and insect pests. It not only seriously threatens our country's forestry resources and ecological construction, but also brings great losses to the national economy. Therefore, the recognition and monitoring of pine trees in the pine forest, the prevention of pine tree diseases and pests, and the protection of pine forest capital are studied. The source and the promotion of sustainable economic and social development are of great significance.
The traditional disease pine tree recognition mainly depends on the artificial field investigation of the disease and insect pest experts, and is affected by the terrain conditions of the mountain area, the responsibility of the investigators and so on. This method is difficult to grasp the epidemic situation in time and accurately, and is difficult to extend in a large scope. With the rapid development of remote sensing technology, the high resolution remote sensing data is obtained by satellite monitoring platform. Monitoring large area of dead pine trees is very effective. However, by satellite operation cycle and atmospheric environment, satellite remote sensing technology is difficult to find pine trees infected by single plant, and it is not suitable for long term application. Therefore, it has become an important research to find an economic and effective way of identification and monitoring. Subject.
In this paper, a small unmanned aerial vehicle (UAV) is used as an aerial flight platform and a dual spectral camera is used as a remote sensing method to obtain the visible and near infrared images of the pine forest. The recognition and monitoring of the pine tree is realized by using image processing, pattern recognition and so on. The main contents are as follows:
1. the purpose and significance of identification and monitoring of pine tree pests and diseases, research status at home and abroad, research areas and data sources are introduced, and the basic theory and development status of support vector data description are introduced.
2. a method of recognition of pine trees for remote sensing images based on weighted support vector data description is proposed. Firstly, according to the characteristics of different objects, the color features of the visible and near infrared images are extracted as the color features of the corresponding pixels, and then the striation of the center pixels is obtained by extracting the gray level co-occurrence matrix of the window image blocks. Then, the weight function of the center distance of the training samples is built to describe the multi classification model of weighted support vector data, and then the multiple output of the ground objects in the image is classified so as to realize the effective recognition of the disease pine trees.
3. a multi classification method of disease pine tree recognition based on multiple features and improved weighted support vector data description is proposed. The method is based on the feature extraction of the single pixel and central local region of each class of samples, and the corresponding feature vectors are constructed. On the basis of the traditional support vector data description multi classification method, the classification model will be in the classification model. The sample of the fuzzy region uses the K nearest neighbor membership function estimation method to make a decision. Combining the advantages of the weight function and the wavelet kernel function, the weighted wavelet support vector data is constructed to describe multiple classifiers and applied to the recognition of the disease pine trees. The experimental results show that the method avoids the traditional multi classification method of the traditional support vector data description. In the fuzzy region, the blindness of the sample decision is to find the sample with high density in the local samples, which greatly improves the recognition rate of the disease pine.
4. a multistage disease recognition method based on feature sparse representation and weighted wavelet support vector data description is proposed. First, the pine tree double spectral image is extracted, then the learning dictionary is constructed by using the feature data set, and the sparse coefficients of each sample are calculated according to the learning dictionary, and the sparse representation feature of the sample is obtained. The weight and input weighted wavelet support vector data are used to describe the multi classifier model, which can effectively classify the pine trees with different degrees of disease. The experimental results show the feasibility and effectiveness of the method in both visual 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

【相似文献】