建筑物作为遥感影像中的主要提取目标之一，其提取具有重要的理论和实际意义。建筑物是制图、土地管理、城市建设规划和灾害评估的重要信息，也是更新地理信息数据库的重要内容。但目前人工影像解译仍然是主要采用的建筑物提取方式。面对海量的遥感数据，人工的方式耗费大量人力和时间，给高分辨率遥感影像的大规模应用带来困难；因此影像中建筑物目标的自动提取成为研究的热点，很多遥感影像建筑物提取方法被提出。随着影像分辨率的提高，传统的基于像素的影像处理方法已不再适用，而地理学的基于对象影像分析方法逐渐成为高分辨率遥感影像处理普遍采用的方式；然而采用该方式的方法往往具有多尺度分割算法中分割尺度的选择问题，而且提取准确度和方法的自动化程度有待提高。

为解决以上问题，本文提出一种基于对象的高分辨率遥感影像建筑物提取方法。方法首先对影像的多尺度分割过程，构造一棵对象树，作为分割的结果。该树记录了分割过程中产生的对象区域和它们间的层次化关系，包含了影像中连续尺度的建筑物对象。再从树上进行对象的特征提取、分类、对建筑物对象的语义分析和结点的筛选，最终提取出建筑物对象。该方法使得提取结果更符合对象实际的尺度，解决了分割尺度的选择问题，提高了提取结果的准确性。

此外，方法对建筑物的特征进行实验和选择，使得选取的特征能很好地区分建筑物和其他目标。方法中的建筑物特征集合融合了多源的信息，既采用了经典的光谱、高度、纹理、几何特征，又提出了一种建筑物的结构特征。方法还采用Adaboost监督学习方法，该方法既能进行学习和分类，又能对特征进行选择和排序，提高方法的鲁棒性和自动化程度。
通过对不同分辨率影像的实验，并和传统方法进行对比，证明了该方法能有效提取影像中的建筑物目标。

It is of great theoretical and practical significance to extract the buildings from the remote sensing images. As the primary part of the remote sensing image of urban areas, buildings are the key information for mapping, land management, urban planning, disaster relief and updating the Geographic Information Database. At present, much of the building extraction is performed by human experts. To process the large amount of images, the way of extraction by human is time consuming and costly, which makes the large-scale application of high resolution remote sensing image difficult. Therefore, extraction of buildings from remote sensing images automatically became a research hotspot and many building extraction methods have been proposed. As the increase of the spatial resolution of the remote sensing images, the traditional pixel-based extraction methods become inadequate for this task, and the methods based on Geographic Object-Based Image Analysis (GEOBIA) are widely used. As to the extraction methods based on GEOBIA, most of them suffer from the problem of the selection of the best segmentation scale and low recognition rate, and cannot work automatically.

In order to solve the above problems of the GEOBIA based methods, this paper proposes a new GEOBIA based method to extract buildings from high resolution remote sensing images. During the multi-scale segmentation procedure of the image, this paper proposes a data structure called object tree, and sets this tree as the multi-scale segmentation results. This tree records the regions created by the segmentation algorithm as object nodes with their hierarchical relationships, and thus contain the building objects with continuous scales. After constructing this object tree, we perform feature extraction, classification and contextual analysis operations on the objects of the tree, and finally select the building objects from these objects as the extraction results. This method can select the building objects with their best scale, which solve the problem of selection of the best segmentation scale, and thus improve the accuracy of the results.

This paper also tests and selects the features of buildings to classify the buildings from other objects. The features used here integrate multi-source information which includes spectral features, height of the objects, textural features and geometry features. Besides, this paper also proposes a structural feature for building extraction. The method also integrates the Adaboost classifier which can not only perform the classification operation automatically, but also select and sort the features.

This paper conducts experiments on images with different spatial resolutions, and compares the results of this method with results of the traditional method. The results demonstrate that this method can extract buildings from the images effectively.