Fusing Multiple Land Cover Products Based on Locally Estimated Map-Reference Cover Type Transition Probabilities

There are a variety of land cover products generated from remote-sensing images. However, misclassification errors in individual and inconsistency among them undermine their utilities for research other applications. While it is worth developing advanced pattern classifiers utilizing the images finer spatial, temporal, and/or spectral resolution increased classification accuracy, also sensible to increase map accuracy through effective fusion by exploiting complementarity multi-source over study area. This paper presents novel method that works weighting multiple source based on map-reference type transition probabilities, which predicted using random forest pixels. The proposed was tested compared with three alternatives: consensus-based weighting, forest, locally modified Dempster–Shafer evidential reasoning, case study, Shaanxi province, China. For this types (GlobeLand30, FROM-GLC, GLC_FCS30) two nominal years (2010 2020) were used as base maps fusion. Reference sample data model training testing collected following robust stratified sampling design allows augmenting reference flexibly. Accuracy assessments show overall accuracies (OAs) fused have been improved (1~9% OAs), outperforming methods 2~8% OAs. does not need products’ systems harmonized beforehand, thus being highly recommendable fusing products.