基于不确定性分析的遥感分类空间分层及评估方法

空间分层是准确度量遥感分类不确定性程度及其空间分布的基础与关键。本文提出了一种基于不确定性分析的遥感分类空间分层及评估方法，首先基于随机森林算法获取像元后验概率，确定分类不确定性度量指标；其次，采用模糊C均值进行空间分层；最后，对分层结果合理性进行定性与定量评估，并与同尺度数据产品精度评价结果及后验概率不确定性分层方法进行对比分析。以北京市顺义区Landsat 8 OLI遥感影像数据为例，研究结果表明：(1)基于最大概率、模糊混淆指数和概率熵指标将顺义区分为不确定性大、中、小3层，相应的遥感数据层分类精度分别为62.28%、74.96%、79.31%；(2)分类不确定性空间分层结果与度量指标大小的空间分布基本一致，错分地类图层与不确定性大层的地类空间分布基本一致；(3)遥感数据和数据产品的各层地类空间特征、层分类精度大小趋势一致，与总体分类精度相比，不确定性大层的层分类精度降低，不确定性小层的层分类精度提高；(4)与后验概率不确定性分层方法相比，本研究不确定性大层的层分类精度降低1.08%，不确定性中层提高3.58%，不确定性小层提高0.16%,q值由0.19提高到0.24，空间分异性...

Spatial Stratification and Evaluation Method of Remote Sensing Classification based on Uncertainty Analysis

Spatial stratification is the basis and key to accurately measure the uncertainty degree and spatial distribution of remote sensing classification. In this study, a spatial stratification and evaluation method of remote sensing classification was proposed based on uncertainty analysis. Firstly, the pixel posterior probability was obtained based on random forest algorithm, and measurement indices of classification uncertainty were determined. Secondly, spatial stratification was achieved using Fuzzy C-means. Finally, the rationality of the stratification results was qualitatively and quantitatively evaluated and compared with the stratification accuracy assessment results of data products at the same scale and the posterior probability uncertainty stratification method. Taking Landsat 8 OLI images of remote sensing data in Shunyi District of Beijing as an example, the results showed that: (1) The coverage of Shunyi District was divided into three strata including large, medium, and small uncertainties based on the indices of maximum probability, fuzzy confusion index, and probability entropy. The stratum classification accuracy of each stratum of remote sensing data was 62.28%, 74.96%, and 79.31%, respectively; (2) 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 misclassified stratum was basically consistent with that of large uncertainty stratum; (3) The spatial characteristics and classification accuracy of each stratum for remote sensing data and data products had the same trend. Compared with the overall accuracy of classification, the stratum classification accuracy of the large uncertainty stratum was reduced, and the stratum classification accuracy of the small uncertainty stratum was improved; (4) Compared with the posterior probability uncertainty stratification method, the stratum classification accuracy of the large uncertainty stratum decreased by 1.08%, and those of the middle uncertainty stratum and the small uncertainty stratum increased by 3.58% and 0.16%, respectively, and corresponding q value of the spatial method developed in this study increased from 0.19 to 0.24 with a higher spatial differentiation. This confirmed that the spatial stratification result of remote sensing classification developed in this study was reasonable. This study proposes an uncertainty stratification scheme for remote sensing classification to optimize the spatial layout of training samples for remote sensing classification and validation samples for accuracy evaluation.