本文以浙江省中南部地区不同时相30m分辨率的2景TM影像为基本数据，采用面向对象的方法实现了研究区的土地覆被制图。首先，在eCognition软件中采用多尺度分割算法，以光谱信息、纹理特征、几何特征等实现研究区的对象分割，使分割后的对象边界与实际地物边界尽量保持一致，通过建立多层次地物特征规则，进行最优分割尺度下的遥感多层次识别分类；然后，分析可用于分类的属性特征和形态特征，通过对这些特征的统计值对比分析，选取了对象的紧致度、长宽比、MNDWI、LBV等特征构建了决策树模型，实现了研究区1：25万的土地覆被分类；最后，采用目视解译和野外样本2种方式对分类结果进行精度验证，其中，目测随机样点评价得到的总体精度为87.66%，野外样本点评价得到的总体精度为83.38%。研究表明：面向对象的分类方法不仅具有较高的精度，而且图斑与实际地物边界能较好地吻合，很好地避免了混合像元误分的现象，同时能消除像元分类的“椒盐现象”。

The land cover map of 30m resolution is generated based on object-oriented method using multi-temporal TM/ETM+ images of central-southern part of Zhejiang Province. The technical process is divided into the following steps. First, multi-scale segmentation algorithm using spectral information, texture characteristics and geometric features is employed to the images of the study area, making the object boundary after segmentation consist to the actual terrain boundaries as far as possible. And by establishing the multi-level object feature roles, we can get different types of the land use with their own extraction scales. This paper uses three-layer split system, the first for parent objects such as woodland and farmland, the second for child objects such as evergreen forest and deciduous forest, and the third for smaller objects such as evergreen coniferous forest and deciduous broadleaved forest. Then, through the analysis of these statistic characteristics, attribute characteristics of MNDWI, LBV and morphological characteristics of compactness, aspect ratio which can be used in classification are analysed, and a decision tree model is constructed to implement the 1:2500 00 land cover mapping of the study area. At last, the precision test of the results are made using two methods of visual interpretation and field validation, and the overall accuracy of visual measurement is 87.66% and the precision of field validating is 83.38%. This article focuses on integration of decision tree algorithm, multi-scale segmentation techniques, hierarchical classification and object-oriented classification method. The results show that the classification method based on object-oriented method not only has high precision, but also realizes the boundaries' coinciding of graph spot and practical ground objects and limits the phenomenon of the wrong classification to the mixed pixels very well. It can also eliminate "pepper phenomenon" based on pixel classification.