Spatial Layout Optimization Of Sample Points For Remote Sensing Classification Accuracy Evaluation Based On Uncertainty Analysis

Reliable remote sensing classification map is the key to analyze land use change.The uncertainty of remote sensing classification makes it necessary to evaluate the accuracy of remote sensing classification map objectively and reliably before use.Sampling method is an important factor affecting the accuracy of accuracy evaluation,simple random sampling and systematic sampling ignore the spatial representation,spatial autocorrelation and heterogeneity of sample points,and there are some problems such as low sampling efficiency and redundant sample information;However,stratified uniform sampling takes into account the spatial representativeness and spatial heterogeneity of sample points,which is closer to the real evaluation of land use types.This paper takes Shunyi District of Beijing as the research area.Firstly,the posterior probability of pixels belonging to each land class is obtained based on random forest algorithm,and the uncertainty measurement index of pixel classification is selected;Then,fuzzy C-means clustering was used for spatial stratification,the optimal number of stratification was calculated,and the rationality of stratification results was evaluated from qualitative and quantitative aspects;Finally,stratified uniform sampling of uncertainty space is constructed in feature space and geographical space,which is compared with other sampling methods.The main conclusions are as follows:(1)Uncertainty measurement indexs are selected and spatial stratification is carried out.The uncertainty measurement indexes of pixel classification are selected: maximum probability,fuzzy confusion index and probability entropy,and the optimal number of layers is determined to be 3.Accord to that degree of uncertainty of pixel classification,Shunyi district is divided into three layer with large uncertainty,medium uncertainty and small uncertainty by fuzzy C-means clustering method.(2)The rationality of spatial stratification was evaluated from qualitative and quantitative perspectives.The study area was divided into three strata including large,medium and small uncertainties.The stratum classification accuracy of each stratum of remote sensing data was62.28%?74.96%and 79.31%,respectively,and the stratum classification accuracy of each stratum of data product at the same scale was 60.86%?63.66% and 65.89% respectively.The spatial stratification results of classification uncertainty were basically consistent with the spatial distribution of the size of measurement indices,and the spatial distribution of type misclassified stratum was basically consistent with that of large uncertainty stratum.The spatial stratification results had the same trend with the spatial characteristics of each stratum and misclassified class strata,and the stratum classification accuracy of the data product at the same scale.It proves the rationality and scientificity of the developed spatial stratification method of remote sensing classification uncertainty.(3)Distribution of feature space and optimal layout of geographical space are carried out for sample points,and stratified uniform sampling of uncertainty space is completed.The number of sample points is determined to be 98,and five groups of data sets with sample points of 98,196,294,392 and 490 are set,and the weight objective function is constructed to complete the feature space allocation of sample points.The optimal layout of geographical space for sample points was realized by using the minimization of the mean of the shortest distances combined with spatial simulated annealing algorithm,Finally,it is integrated into the stratified uniform sampling of uncertainty space.(4)The accuracy evaluation index system of overall classification accuracy,relative accuracy,root mean square error and standard deviation is constructed to evaluate the accuracy of sampling methods.For Shunyi study area,the mean and standard deviation of the overall classification accuracy of the five sample point data sets of uncertainty spatial stratified uniform sampling,simple random sampling,stratified random sampling and spatial uniform sampling methods are 0.716±0.036 ? 0.737±0.036 ? 0.715±0.036 ? 0.737±0.032 respectively,and the corresponding root mean square errors are 0.033,0.037,0.037 and 0.037respectively;For the large layer of uncertainty,the mean and standard deviation of the overall classification accuracy of the five groups of sample point data sets of stratified uniform sampling and stratified random sampling methods are 0.625 ± 0.040 and 0.618 ± 0.060 respectively,and the corresponding root mean square errors are 0.036 and 0.076 respectively.The comprehensive analysis shows that the uncertainty space stratified uniform sampling method has better sampling effect than other sampling methods.(5)The uncertainty space stratified uniform sampling method proposed in this study not only improves the uniformity of sampling points in geographical space,but also improves the representativeness of sampling points in feature space.The purpose of this research is to provide a way of thinking and reference for remote sensing classification accuracy evaluation and authenticity testing.However,there are still some uncertainties in the selection of training samples and indexs.