Fusing Multiple Land Cover Products Based On Locally Estimated Map-reference Cover Type Transition Probabilities

With the development of satellite remote sensing,the research on land cover mapping based on remote sensing images has been carried out one after another,forming a number of regional or global scale land cover remote sensing products.Due to the differences in classification algorithms and category systems in the mapping process,there are numerous problems in land cover products,such as thematic errors and insufficient consistency between products.The fusion of multi-source land cover data provides a way to solve the above problems.Aiming at the problems of current land cover products and the shortcomings of existing fusion algorithms,this paper proposes the concept of local map-reference cover type transition probabilities and a consensus fusion method based on local type transition probabilities.Shaanxi Province was selected as the research area,and three types of land cover products(Globeland30,GLC＿FCS30 and FROM-GLC)of two nominal years of 2010 and 2020 were used as base maps for fusion.It works by weighting multiple source products based on their map-reference cover type transition probabilities,which are predicted using random forest from map class to reference class for individual map pixels,and compared with the conventional consensus map fusion method.Secondly,the method based on random forest and DempsterShafer evidence reasoning are used to fuse multi-source land cover products.Finally,the accuracy of the base land cover products and the fusion products was verified according to the verification samples extracted from the supplementary sampling design.The results show that the consensus fusion method based on local type transition probabilities achieves the maximum overall accuracy gain,and the accuracy is improved by about 6.32%-9.29%.The DempsterShafer evidence reasoning method and the random forest method are second,and the accuracy is improved by about 2.62%-7.34%.The consensus map fusion method has the worst accuracy improvement effect,about-0.05%-1.04%.The difference between the consensus fusion method based on local transfer probability and the consensus map fusion method shows the great advantages of local type transition probabilities in the process of multi-source land cover information fusion.In addition,in the process of estimating the local type transition probabilities,it is only necessary to specify a reference classification system to estimate the local type transition probabilities from the classification system of any product to the reference classification system,so as to realize the fusion of multi-source land cover product.Compared with other fusion methods,the consensus fusion method based on local type transition probabilities is not limited by the base land cover product classification system.