论文部分内容阅读
综合运用自然、生态和人文等多元知识辅助计算机遥感自动分类,确保大区域地表覆盖遥感数据产品质量是多年来国际上致力于解决的科技难题.在全球30m地表覆盖遥感数据产品GlobeLand30的研制过程中,作者提出了一种顾及多元知识的检核优化模型,对基于“像元-对象”的遥感自动分类结果进行知识化检核,发现和修改错分/漏分问题.首先归纳总结全球地表覆盖分布相关的先验知识,凝练出有关地表覆盖的地域地形分布规律、时间连续性和空间关系等质量检核知识规则;继而发展了基于网络化服务平台的交互式协同检核技术,辅助发现与定位分类异常区域,并依据各类别最小图斑规定判定错分/漏分,进而标注和发布错误信息,供研制人员用于修改优化.以林灌草和裸地数据为例,抽取10%的图幅,进行优化前后的分类精度对比分析,得知质量明显改善,精度提高最多达到17.47%.运用该模型,对GlobeLand30数据进行了逐类检核与优化处理,有效地减少了错漏分问题,2010期产品的总体精度达到83.50%.这一模型还可用于其他全球或区域的地表覆盖数据产品研制.
Comprehensive application of multiple knowledge of nature, ecology and humanities to assist in the automatic classification of computer remote sensing to ensure the quality of the remote sensing data of the land cover in large area is a scientific and technological problem that has been devoted to solve for many years in the world.In the process of developing globe 30m remote sensing data product GlobeLand30 , The author puts forward a checking and optimizing model that takes into account the multiple knowledge to carry out the knowledge check on the results of the automatic classification of remote sensing based on the “pixel-object” to find and modify the misclassification / omission problem.Firstly, The priori knowledge related to the distribution of surface cover, and concludes the knowledge rules about the quality inspection of geographical coverage, time continuity and spatial relations; then develops the interactive collaborative checking technology based on networked service platform to assist Find and locate the classification of anomalous areas, and determine the classification according to the provisions of the smallest spots of each category misclassification / leakage points, and then annotate and release error information for researchers to modify and optimize .In the forest shrubs and bare land data, for example, extract 10 % Of the map, before and after the optimization of the classification accuracy comparative analysis, that the quality was significantly improved, the accuracy of up to 17.47 %. Using this model, GlobeLand30 data is checked and optimized by class by class, which can effectively reduce the number of errors and omissions. The overall accuracy of the 2010 product is 83.50% .This model can also be used for other global or regional land cover Data product development.