面向对象图像识别的地学理解方法的优化与应用

近年来,面向对象技术的应用愈加广泛,以适应不同应用目的下不同地物复杂度的图像处理与信息提取要求。本文详细分析了面向对象信息提取方法中存在的问题,提出将传统像元级光谱分类与面向对象信息提取技术相结合,通过最大似然光谱分类和地学统计分析获取图像区域的先验知识,包括不同类型地物光谱混淆情况和图像复杂程度的定性描述等,以反映图像处理与信息提取的难度大小及技术要求,利用预分析获得的知识优化面向对象信息提取技术流程中图像分割参数设置及分类器的构造等,用于克服面向对象方法参数选取的盲目性和结果的多样性。并以SPOT5高分辨率遥感图像为例,通过常规方法和改进方法的对比,证明本试验方法有利于真实地分割地物基元,提高分类精度、效率和训练器性能。

Optimization and Application of Object-oriented Information Extraction Methods Based on Geographic Understanding of Remote Sensing Images

In the academic domain of remote sensing information extraction,per-pixel spectral classification and object-oriented classification method are two main approaches.Presently,the application scope of object-oriented classification method is becoming enlarged so as to meet various needs of image processing and information extraction towards diverse ground objectives.The object-oriented method has been a useful helper in many cases;unfortunately it has not developed into a great approach.In fact,several problems have been spotted when applying object-oriented methods.This paper intends to deal with the problems in executing object-oriented information extraction and proposes a new concept,which is to combine the per-pixel classification and object-oriented classification,based on above mentioned delicate problem-studying.The methodology of this paper is presented as follows.To begin with,the priori regional knowledge is first obtained by means of geographic analysis based on traditional per-pixel maximum likelihood classification.By doing this,the spectral confusion between different types of ground objects and the complexity of the whole image can be described both quantitatively and qualitatively,in order to reflect the difficulty and technological requirements regarding image processing and information extraction.Subsequently,the knowledge gained via pre-analysis is then used to direct the technique taches in image processing,as well as to optimize the segmentation parameters and the classifier construction in the workflow of object-oriented extraction.A major objective of the abovementioned technique is to conquer the blindness of parameters setting and the multiformity of segmentation results.Finally,compared to the object-oriented result without the support from geographic understanding and analysis,this proposed method is proved good at partitioning the real elementary objects and improving the classification accuracy,efficiency and the training performance.