Sensitivity Analysis of Pre-classification Accuracy Based on Remote Sensing Image to Ground Area Estimation from Spatial Sampling
遥感影像预分类精度对地物面积空间抽样估算的敏感性分析

The traditional ground areameasurement is lmi ited by accuracy and efficiency. Remote sensing technology helps mi prove both butalone still far from satisfaction. Therefore, we proposed a spatial stratified samplingmethod based on remote sensing. The basic idea is using up to date remote sensing mi ages to guide the target stratified sampling by a general classification, other than by those obsolete background knowledge databases. In order to validate the mi provement on accuracy ofcalculating the realquantitywith actually itunknown (which is justourgoal to estmi ate), we took the early classificationmap from the remote sensing mi age ( in this expermi ent, we used partofone scene ofTM mi age coveringBei- jing area, a size of4800×4800 pixels, and classified into 5 different types, 3 ofwhichwere chosen)as the laboratorial quasi-real target (proportions vary from 7% to 32% ). The detailmethods and operations are described in the following steps: Firstly, we played back smi ulated pre-classi- fication result, which completely contains the target, by iteratively adding error classpixels around the outskirtsof the tar- get to a demanded proportion; secondly. Secondly, we arranged square boxes (with a size of20×20 pixels)on the pre- classification mi age, excluded zero-targetones and divided the rest into 5 strata according to the proportion of in-box pre- classification targetarea (pixels), randomly orsystematically chose the samples. And thenwe estmi ated the grossby sum- ing up the actual pixels in each sample pro rata. Finally, we analyzed and compared the quality and variation ofestmi a- tion accuracy repetitiouslywith different land cover types, differentpre-classification precision levels, and twomethods of random and system in stratum. The resultsmainly presented the relationship between estmi ation accuracy and pre-classification accuracy in each tar- get type, which showed that the estmi ation accuracy degraded when the stratified samplingmethod was aided with rough pre-classifications (accuracy less than 40% ), but remarkably reached higher accuracy and stabilization than those ofun- supported random or systematic samplingmethodswith pre-classification above a certain accuracy leve.l For the former sit- uation, the degradation ismainly caused by the extreme inconsistency ofarea distributing direction poorly classified, and does not take place in common classifications. In genera,l thismethod has itsbestcost-efficiency at50% pre-classification accuracy, a case ofwhich is that the ac- curacy of the estmi ators for each target classwith all proportions at0.5% sampling ratio level and 95% confidence level can be acquired above 92% when the accuracy ofpre-classification reaches60%. During the study, we innovated in the following aspects: first, we created a smi ulated classification map with an assumed given target, and thismap worked properly to show up the real situation; second, with the help ofpre-classification from remote sensing mi ages, t