Spatial stratification method for the sampling design of remote sensing classification accuracy assessment

Spatial sampling design is essential for accurately assessing land use and land cover (LULC) classification results from remote sensing data. When classification correctness exhibits spatial heterogeneity, spatial stratification can significantly improve spatial sampling efficiency by dividing the study area into heterogeneous strata. Three spatial stratification methods were introduced, respectively focusing on LULC types, the integration of multi-source classification products with different spatial resolutions, and pixel-level uncertainty analysis.

First, stratification by LULC types was employed because these categories directly relate to variations in classification accuracy. Second, although LULC products from different sources and resolutions were generated using diverse data and methods, their consistency and inconsistency could indicate potential misclassification. Thus, a stratification method that combined such multi-source products was developed for guiding accuracy assessment sampling. Third, a pixel-based stratification framework was proposed based on uncertainty indices, namely the maximum probability, fuzzy confusion index, and probability entropy.

The effectiveness of these methods was tested through a case study of LULC classification in Beijing, China. Results showed that the proposed stratification approaches could effectively distinguish spatial characteristics and improve sample representativeness, thereby optimizing the sampling for classification accuracy evaluation and enhancing its overall reliability.