Abstract: Taken Landsat-5 TM image in Shenyang city, Liaoning province as a data source, ERDAS software was used for supervised classification and eCognition software for object-oriented classification. Multi-scale segmentation method in object-oriented classification and the nearest neighbor classification were selected to classify. After classification, the error matrix of supervised classification was calculated to obtain overall accuracy, kappa coefficient, producer accuracy and user accuracy. The best classification result and classification stability method were selected to assess accuracy in eCognition software. The results were (supervised classification/object-oriented classification): settlement 91.14%/93.50%, farmland 86.91%/93.80%, forest 91.73%/ 96.70%, grassland 84.44%/91.36%,water 98.16%/96.18%. The result of object-oriented classification method improves the efficiency and accuracy of classification compared to traditional supervised classification method.

Key words: land cover, object-oriented, TM image, multi-scale segmentation, nearest neighbor algorithm

摘要： 以辽宁省沈阳市Landsat-5 TM影像为数据源，利用ERDAS和eCognition软件分别使用监督分类和面向对象分类方法对试验区土地覆被进行分类。在面向对象分类中选用多尺度分割法、最 邻近分类法。分类完成后，计算监督分类误差矩阵得到总体精度90.65%、Kappa系数0.847 8以及生产精度和用户精度，在eCognition中选用最佳分类结果和分类稳定性统计法得到图表形式的 精度结果，最后得到精度结果对比(监督分类/面向对象：聚落91.14%/93.50%,农田86.91%/93.80%,森林91.73%/96.70%，草地84.44%/91.36%，水体98.16%/96.18%。通过精度对比分析得出基 于对象的面向对象分类法较于传统的监督分类法提高了分类效率和精度。

关键词: 土地覆被, 面向对象, TM影像, 多尺度分割, 最邻近算法