摘要
首先根据生产者精度和使用者精度的概念,提出生产者精度和使用者精度的条件概率表达式,然后根据概率乘积公式,推导出生产者精度和使用者精度之间的关系式。该关系式表明:①使用者精度和生产者精度的比值可作为类别真实面积与分类结果中类别面积的比值的估计;②利用使用者精度和生产者精度的比值可对遥感分类结果进行修正,产生更接近于真实值的土地覆盖类别面积值,且该方法的计算结果仅取决于使用者精度和生产者精度数据的可靠性,与分类算法的优劣无关。该方法可用于最大似然分类方法中先验概率的估计。对ErdasImagine软件所附带lanier.img文件的实验结果表明,各种分类结果包括一种对常规最大似然分类结果进行任意修改后的分类结果,利用文中提出方法修正后均产生了接近于真实值的类别面积比例。由于作为标准的精度检验方法,几乎所有的分类影像都会产生误差矩阵用于精度报告,这保证了该方法具有很好的应用价值,可以帮助土地利用/土地覆盖研究中获取更准确的土地利用/土地覆盖的面积数据。
Producer＇s accuracy and user＇s accuracy, which are usually derived from error matrix, are often used to measure individual class accuracy. Producer＇s accuracy indicates the probability of a reference pixel being correctly classified. User＇ s accuracy is indicative of the probability that a pixel classified on the map actually represents that category on the ground. Based on the concept of producer＇ s accuracy and user＇s accuracy, their conditional probability expressions are proposed and their relationship is defined according to the probability-product rule. Their relationship implies ： 1 ） the ratio of producer＇s accuracy to user＇s accuracy is equal to the ratio of the true class area to the class area in output classification, and then it can be used as one indicator of the accuracy of the class area in output classification： the closer to unity it is, the closer to true value the class area in output classification. 2） The ratio of producer＇ s accuracy and user＇ s accuracy can be used to revise the class area obtained by direct classification procedure, and after this revision, the class area much closer to true value will be obtained. And this method performance, only determined by accuracy of the data of producer＇s accuracy to user＇s accuracy, is independent on whether the classifier is optimal or problematic. So, this method can be used to estimate prior probabilities in maximum likelihood classification, which are relative area proportions for individual classes in remote sensing image to be classified.
Two case studies are presented. One involved in the Landsat TM imagery covering Twente region, The Netherlands. The other involved in lanier, img file attached to Erdas Image software. The result shows： the ratio of producer＇s accuracy to user＇s accuracy can be used as a good indictor for the estimation accuracy of class area, and whether this ratio is close to unity is generally consistent with the variation of overall accuracy or kappa index. For all of five classification results involving in lanier, img file, one of which is generated by random revision of the output of maximum likelihood classification using equal priors, the proposed method produces the class area estimation closer to reality than in the output classification.
Since the error matrix, the standard method of accuracy assessment of classifications of remotely sensed data, will be produced after the classification for almost all the classification procedures. This method will be helpful to obtain better class area estimation and further to improve classification accuracy of maximum likelihood classifier.
出处
《资源科学》
CSSCI
CSCD
北大核心
2007年第3期214-220,共7页
Resources Science
基金
国家基金委员会重点基金项目(编号:40635027)
国家自然科学基金资助项目(编号:40271075)
中科院地理科学与资源研究所知识创新工程项目(编号:CXIOG-A04-10)
关键词
土地覆盖
面积估计
先验概率
最大似然法
Land cover
Class area estimation
Prior probability
Maximum likelihood classification
