Spatiotemporal consistency correction of multi-temporal GlobeLand30 land-cover products using a texture-informed MAP–MRF model

Multi-temporal GlobeLand30 products provide 30 m land-cover information with generally high classification accuracy, but residual spatial inconsistency and illogical temporal transitions can limit their use in land-cover change analysis. This study proposes a MAP–MRF-based post-processing method to correct inconsistencies in the 2000, 2010, and 2020 GlobeLand30 products. Because the released classification maps no longer contain spectral observations, texture features were used to describe the observation field, while spatial neighbourhoods, temporal neighbourhoods, and transition-rationality constraints were integrated into the MAP–MRF framework. The processed products were evaluated using spatial consistency, temporal consistency, and classification accuracy. Results from the Jiangsu study area show that the original products already had high spatial consistency. The spatial MAP–MRF mainly improved local spatial coherence, whereas the spatiotemporal MAP–MRF reduced illogical temporal transitions, while the mean classification accuracy showed a modest numerical increase from 82.0% to 83.5%. Overall, the proposed post-processing method improves the temporal reliability of three-phase GlobeLand30 products with limited disturbance to the original classification structure.