高分辨率遥感影像分割方法及应用研究

发布时间：2018-03-06 08:35

本文选题：高空间分辨率遥感影像　切入点：影像分割　出处：《长安大学》2016年博士论文　论文类型：学位论文

【摘要】：随着遥感对地观测技术的不断进步,遥感影像的空间分辨率越来越高,高空间分辨率影像为遥感技术的发展带来机遇的同时,也由于自身的特点为遥感数据的处理带来了新的挑战,导致传统的像素级处理方法不能适应当前遥感分析与应用的需求,在这种情况下,面向对象技术成为高分辨率遥感影像分析的新选择。然而,面向对象分析在理论和技术上并未达到成熟,仍然有许多问题亟待研究和解决,突出表现在遥感影像分割这一基础环节,目前已有的研究仍然存在不同程度的局限性,如分割精度与效率的不足、模型的适应性差、缺乏统一的尺度评价标准等。本文针对现有研究工作中的不足,重点对高空间分辨率遥感影像的分割及多尺度分割中的最优分割尺度选择问题进行了探索与讨论,主要研究内容与成果如下:(1)提出了一种基于区域合并的遥感影像快速分割方法。该方法首先通过分水岭变换得到影像的初始分割区域,然后采用局部寻优的策略对初分割区域进行合并得到最终分割结果。在区域合并过程中,针对传统的区域邻接图效率较低的问题,设计了邻域数组的数据结构来维护区域之间的邻接关系,该数据结构相比邻接图具有更高的效率,为兼顾算法的精度,方法将区域合并过程分为两个步骤进行,在合并准则中综合考虑了区域的光谱、几何和纹理信息。通过与传统的基于邻接图的区域类分割算法的对比,证明了算法在分割效率和精度上的优势。(2)提出了一种融合影像多种特征的多尺度分割方法。该方法也属于区域类的分割算法,在区域合并时综合利用影像的光谱、纹理、形状等特征来构建对象的语义描述并建立合并规则,以解决影像特征利用不足而导致分割精度较低的问题,算法引入非下采样轮廓波变换并结合模糊聚类分析来计算对象的纹理特征距离。在合并过程中采用全局寻优的策略保证了方法的多尺度特性,并联合采用邻接图和最优邻图两种图模型来维护区域的邻接关系,保证了算法的效率。通过与其它算法的对比,证明了算法在精度和效率上已可匹敌同类商业软件。(3)提出了一种遥感影像最优分割尺度的监督评价方法。该方法根据参考对象和分割对象的相似性来构建分割尺度评价函数,通过计算评价函数值来确定最优分割尺度。采用统计直方图来进行灰度相似性的计算,克服了其它特征进行对象灰度描述时的不准确性,通过构建形状描述函数进行形状相似性计算,能够对对象的几何差异作出较为准确的判断。通过实验将该方法与较为成熟的人为试错法相比,验证了算法的有效性。(4)发展了一种基于面向对象的高分辨率遥感影像道路提取方法。该方法首先应用本文提出的分割算法及最优分割尺度评价方法得到道路影像在最优尺度下的分割结果,然后通过构建道路知识库,并进行地物特征的计算,来实现道路的初步提取;最后对道路初始提取结果进行多方向形态学滤波,以去除特征相似的混叠地物,优化道路提取结果,并对优化后结果进行细化及连接处理,从而实现道路网的提取。实验结果表明,该方法能较好地从复杂遥感影像中提取道路网。
[Abstract]:With the remote sensing of earth observation technology continues to progress, the spatial resolution of remote sensing images more and more high, the high spatial resolution image brings opportunities for the development of remote sensing technology at the same time, but also because of their own characteristics for the processing of remote sensing data has brought new challenges, leading to the traditional pixel level processing method can not adapt to the current remote sensing analysis and application the demand, in this case, the object oriented technology has become a new choice of high resolution remote sensing images. However, object oriented analysis has not reached maturity in theory and technology, there are still many problems to solve, especially in the remote sensing image segmentation is a basic step, the existing research still has limitations the degree, such as the lack of segmentation accuracy and efficiency, the adaptability of the model is poor, lack of unified evaluation standard scale. Based on the existing research work in Lack of focus on the optimal segmentation of high spatial resolution remote sensing image segmentation and multi-scale segmentation scale selection problems are discussed. The main research contents and results are as follows: (1) proposed a fast segmentation method of remote sensing image based on region merging. This method firstly by the initial watershed transform to obtain image segmentation region, and then uses a local optimization strategy of early segmentation region merging to get the segmentation result. In the process of region merging, the traditional region adjacency graph and the problem of low efficiency, the data structure is designed to maintain the number of neighborhood group adjacency relation between regions, compared to the data structure graph is more efficient for both, the accuracy of the algorithm, the method of region merging process is divided into two steps, in the region merging criterion in spectra considering geometric and texture information. With the comparison of adjacency graph area segmentation algorithm based on the traditional algorithm, show the advantage in the segmentation efficiency and accuracy. (2) proposed a segmentation method based on multi-scale fusion multi feature of image segmentation algorithm. This method also belongs to the area, in the area with comprehensive use of image spectrum. The texture, shape and other characteristics to construct the object semantic description and establish merger rules to solve the shortage of image features using the low precision problem of segmentation algorithm is introduced, Nonsubsampled contourlet transform and fuzzy clustering analysis to calculate the distance of texture feature images. In the process of merging with global optimization strategy the multi-scale characteristic method, and combined with the adjacency graph and the optimal neighborhood graph two graph model to maintain relationship between adjacent areas, to ensure the efficiency of the algorithm. By comparing with other algorithms, proved The algorithm in accuracy and efficiency have been unmatched similar commercial software. (3) proposed a supervision and evaluation method of remote sensing image optimal segmentation scale. According to the similarity of the reference object and object segmentation to construct the segmentation scale evaluation function, through the calculation to determine the optimal segmentation scale by statistical histogram evaluation function. Calculate the similarity of gray, overcome the other characteristics of the intensity of the object description inaccuracy, by constructing a shape similarity calculation function of describing the shape of the object to the geometrical differences to make more accurate judgments. The experiment will be compared with the mature method of trial and error method, verify the validity of the algorithm. (4) developed a road extraction method of high resolution remote sensing image based on object oriented segmentation algorithm and optimal scale. Firstly, this paper presents the application of The evaluation method of road image segmentation result is obtained in the optimal scale, and then through the construction of road knowledge base, and compute the features extracted, to achieve the road; finally, multi direction morphological filtering for initial road extraction results, similar to mixed stack objects removal characteristics, optimization of road extraction results, and the optimized the results of refinement and connection processing, so as to realize road extraction. The experimental results show that this method can effectively extract the road network from the complex remote sensing image.

【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：P237

【相似文献】