Mangrove forests are swampy, woody plant communities mainly composed of evergreen shrubs or trees distributed in the coastal wetlands of tropical and subtropical regions throughout the world. They have special sea-land characteristics and are a wetland ecosystem with important ecological, economic, and landscape values in the coastal zone. Four periods of global mangrove distribution dataset for 2000, 2010, 2015 and 2020 were generated by mapping the global mangrove change with a combination of the Capsules-Unet deep learning model and visual interpretation of imagery using multi-temporal Landsat images with no clouds or less than 20% cloud cover after pre-processing such as atmospheric correction and band combination. Compared with similar products, this product has a longer time series and can accurately reflect the dynamics of global mangrove changes. The overall accuracy of verification is greater than 88%, kappa coefficient and F1 index are greater than 0.76 and 0.88 respectively.