基于无人机多光谱遥感影像的地物分类方法研究

发布时间：2018-08-19 17:10

【摘要】：遥感影像的获取与分类技术是遥感监测过程中的基础和关键。以往的航天航空遥感虽然能快速获取大面积遥感影像,但是空间、时间分辨率较低。随着当前无人机和轻小型传感器的发展,使高空间、时间分辨率的低空影像的获取成为可能。当前多光谱影像在应用中易产生“同谱异物”、“同物异谱”等现象,更高的光谱分辨率也让相邻波段间的相关性大大增加,使得计算复杂度和时间复杂度大幅提高。因此对多(高)光谱遥感影像数据进行降维处理是目前应用研究中的一个难点问题。本文面向低空多光谱地物分类,基于最佳波段指数和影像的光谱特征、纹理特征进行最佳波段组合的选择,然后利用支持向量机和最小二乘支持向量机分类方法构建了多组分类模型进行分类对比实验。主要工作及相关研究成果如下:(1)利用大型固定翼无人机搭载轻型多光谱相机搭建无人机遥感影像采集平台并获取了地面分辨率为22.6cm的12波段无人机多光谱遥感影像,再通过Pix4D Mapper对原始图像进行基于特征的配准和特征级的融合得到研究区域的正射影像。(2)针对无人机多光谱影像数据的空间分辨率高、波段间相关性大等特点,综合影像的植被及水体指数等光谱信息,影像主成分分析和灰度共生矩阵计算得到的纹理特征信息及最佳波段指数法筛选的原始波段得到了最佳波段组合来进行地物分类。(3)针对研究区初始得到的波段组合,设计监督分类和非监督分类的对比实验。相对原始波段组合,其中研究区域A的1,6,11,NDVI,NDWI,Mean波段组合的IsoData分类精度从83.57%提高到89.80%,SVM分类精度95.58%提高到99.76%。实验证明此波段组合不仅包含较多的波段信息且波段间相关性系数较低,同时反映了地物的光谱信息和纹理信息,可选择其作为Micro MCA12 Snap的最佳波段组合。(4)针对实验得到的最佳波段组合,分别使用粒子群优化和网格搜索算法进行参数寻优并使用交叉验证的方法对研究区域进行SVM和LSSVM对比实验。实验结果表明,以粒子群优化进行参数寻优得到的LSSVM分类模型,相对SVM粒子群优化分类精度从97.833%提高到99.854%;相对LSSVM网格搜索分类精度从99.762%提高到99.854%。同时LSSVM粒子群优化在一定程度上提高了分类的速度,是针对本文最佳波段组合在地物分类上的理想分类模型。
[Abstract]:The acquisition and classification of remote sensing images is the basis and key in the process of remote sensing monitoring. Although space and aviation remote sensing can obtain large area remote sensing image quickly, the spatial and temporal resolution is low. With the development of UAV and light sensor, it is possible to obtain low-altitude images with high spatial and temporal resolution. At present, multi-spectral images are easy to produce such phenomena as "isospectral foreign bodies" and "isospectral spectra". The higher spectral resolution also increases the correlation between adjacent bands, which greatly increases the computational complexity and time complexity. Therefore, dimensionality reduction of multispectral remote sensing image data is a difficult problem in application research. In this paper, we choose the best band combination based on the best spectral index and the spectral feature of the image, and the texture feature is used to select the best band combination for the low altitude multi-spectral feature classification. Then, support vector machine (SVM) and least squares support vector machine (LS-SVM) are used to construct multi-group classification models for classification comparison. The main work and related research results are as follows: (1) using a large fixed-wing UAV with a light multi-spectral camera to build a UAV remote sensing image acquisition platform and obtain 12-band UAV multi-spectral remote sensing image with ground resolution of 22.6cm. Then, the orthophoto image of the study area is obtained by the feature based registration and feature level fusion of the original image by Pix4D Mapper. (2) aiming at the characteristics of the UAV multi-spectral image data, such as high spatial resolution and large correlation between bands, etc. Spectral information such as vegetation and water body index of integrated image, The texture feature information obtained by principal component analysis and gray level co-occurrence matrix calculation and the original wave band selected by the best band index method are selected to obtain the best band combination for ground object classification. (3) for the initial band combination in the study area, Design the contrast experiment between supervised classification and unsupervised classification. Compared with the original band combination, the IsoData classification accuracy of the study area A is improved from 83.57% to 89.80% from 83.57% to 99.76%. The experimental results show that the band combination not only contains more band information, but also reflects the spectral information and texture information. It can be chosen as the best band combination of Micro MCA12 Snap. (4) for the best band combination obtained from the experiment, Particle swarm optimization (PSO) and mesh search algorithm are used to optimize the parameters, and the cross-validation method is used to carry out the SVM and LSSVM comparative experiments in the study area. The experimental results show that the classification accuracy of relative SVM particle swarm optimization is improved from 97.833% to 99.854, and that of relative LSSVM mesh search and classification is improved from 99.762% to 99.854. At the same time, LSSVM particle swarm optimization improves the speed of classification to a certain extent, which is an ideal classification model for the best band combination in this paper.
【学位授予单位】：石河子大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237

【参考文献】