In order to evaluate the sampling accuracy of remote sensing classification, taking Beijing-Tianjin-Hebei remote sensing data products with different spatial resolutions as an example, the internal and boundary objects of remote sensing image were firstly divided based on land use types, and different spatial stratification modes were constructed, including without considering internal and boundary objects, considering boundary objects, considering internal objects, both considering internal and boundary objects. Secondly, direct land use types, image eight-neighborhoods algorithm, multi-scale spatial differentiation method, coupling method of image eight-neighborhoods and multi-scale spatial differentiation were adopted for spatial stratification, respectively. Finally, a comparative experiment of K-means clustering method was set up, and the differentiation effects of different spatial stratification modes were quantitatively evaluated based on geographic detector. The results suggested that the mean and standard deviation of qof the corresponding five groups of sampling sites for the spatial stratification modes of without considering internal and boundary (6 strata), considering boundary (12 strata), considering internal (18 strata), both considering internal and boundary objects (24 strata), K-means (12, 18, 24 strata) in the Beijing-Tianjin-Hebei regions were 0.252±0.02266, 0.259±0.02245, 0.321±0.01901, 0.318±0.01806, 0.269±0.00698, 0.304±0.01056, and 0.317±0.01125, respectively. Internal objects played a leading role for spatial stratification differentiation and boundary objects slightly improved spatial stratification differentiation, and the number of strata also affected the differentiation of spatial stratification. The research results can better understand the contributions of internal and boundary objects on improving spatial stratification differentiation, and had a certain research value and guiding significance for developing spatial stratification methods with high differentiation.