Land cover map can accurately characterize the spatial distribution of natural and artificial surface features.However,large-scale land cover products with submeter resolution are still scarce.To address this gap,this...Land cover map can accurately characterize the spatial distribution of natural and artificial surface features.However,large-scale land cover products with submeter resolution are still scarce.To address this gap,this study proposes an innovative data annotation engine,called initial and expanded labeling,to generate reliable labels for high-resolution imagery.The engine takes imagery and historical products as input,generates a small number of labels using weight voting in the first stage,and iteratively expands the labels in the second stage.The proposed method can effectively deal with the insufficiency of training labels in large-scale submeter land cover mapping.Based on the datasets generated by this engine,we have produced the first large-scale submeter land cover map covering the urban areas of 42 major cities in China,called EcoVision.It has a spatial resolution of about 0.5 m with 8 representative urban land cover categories.The product has been validated with 23,850,000 randomly sampled validation pixels in 42 cities and has an overall accuracy of 83.6%.Compared with 5 existing land cover maps,EcoVision shows superior performance in spatial resolution,accuracy,and details.The product has been made public,providing high-precision data support for urban sustainable development research and territorial spatial planning.展开更多
基金supported by the National Key Research and Development Program of China(grant number 2024YFF1306102)the National Natural Science Foundation of China(grant numbers 42271328 and 42471391).
文摘Land cover map can accurately characterize the spatial distribution of natural and artificial surface features.However,large-scale land cover products with submeter resolution are still scarce.To address this gap,this study proposes an innovative data annotation engine,called initial and expanded labeling,to generate reliable labels for high-resolution imagery.The engine takes imagery and historical products as input,generates a small number of labels using weight voting in the first stage,and iteratively expands the labels in the second stage.The proposed method can effectively deal with the insufficiency of training labels in large-scale submeter land cover mapping.Based on the datasets generated by this engine,we have produced the first large-scale submeter land cover map covering the urban areas of 42 major cities in China,called EcoVision.It has a spatial resolution of about 0.5 m with 8 representative urban land cover categories.The product has been validated with 23,850,000 randomly sampled validation pixels in 42 cities and has an overall accuracy of 83.6%.Compared with 5 existing land cover maps,EcoVision shows superior performance in spatial resolution,accuracy,and details.The product has been made public,providing high-precision data support for urban sustainable development research and territorial spatial planning.
