前期土地覆被数据辅助下的分类样本自动选取

 将地学知识与影像标定相结合,一直是目视解译或计算机自动分类制图的主要手段。传统的目视解译方法能够充分利用地学知识,但需要大量的人力、物力,效率较低;计算机分类中尚未出现比较成熟的高效运用地学知识的分类方法。已有研究表明,分类样本可以作为地学知识的载体,将地学知识融入分类过程中;此外,无监督聚类可以显著提高样本选取的效率,有助于提供足够的样本,为将地学知识高效地融入计算机分类提供了一定的基础。本文提出一种以前期土地利用数据辅助与影像聚类相结合的样本自动选取方法。利用自动选取的样本,通过最大似然分类器对TM影像进行分类,并与手动选取样本分类的方法进行了对比分析。研究结果表明,在分类效果上,本文提出的前期土地覆被辅助下的分类样本自动选取方法,优于手动选取样本的方法,提高了分类效率。在水体、林地、园地、城镇建设用地等7种类型上的分类整体精度达到84.18%,kappa系数为0.8066;手动选取样本进行分类的整体精度为77.04%,kappa系数为0.7196。

Automatic Selection of Classified Samples with the Help of Previous Land Cover Data

The combination of geographical knowledge and image calibration has long been the principal means of both the traditional visual interpretation and computer automatic classification in remote sensing mapping.Traditional visual interpretation could use the geographic knowledge well because of the artificial participation.However,it goes with the shortcomings that visual interpretation needs a lot of labor and is less efficient.In addition,the computer classification has not applied geographic knowledge in a proper way.Studies have shown that samples as the carrier of geographic knowledge can integrate geographic knowledge into the classification process to some extent.Meanwhile,unsupervised clustering can significantly improve the efficiency of sample selection and solve the problem of scarcity of samples in order to meet the requirement of distribution and purity.These studies provide a basic foundation for integration of geographic knowledge with computer classification.This paper presents an automatic sample selecting method which integrates image clustering with the aid of previous land cover data.The samples were selected automatically based on the TM images by the method mentioned above and used to classify the image later by the maximum likelihood classifier.We also classified the image using the manual samples by the maximum likelihood classifier in order to compare the classified results produced by these two kinds of samples.The test results indicated that the proposed method achieved an overall accuracy of 84.18% and a kappa coefficient of 0.8066 in seven categories,including water body,forest land,orchard and urban construction land.The method proposed in this paper is more efficient than the way of samples selected manually and provides better classification results.