Fig. 3
From: Dynamic World, Near real-time global 10 m land use land cover mapping
Training inputs workflow. Annotations created using Sentinel-2 Level 2 A Surface Reflectance imagery are paired with masked and normalized Sentinel-2 Level 1 C Top of Atmosphere imagery, and inputs are augmented to create training inputs used for modeling. Cloud and shadow masking involves a three-step process that combines the Sentinel-2 Cloud Probability (S2C) product with the Cloud Displacement Index (CDI), which is used to correct over-masking of bright non-cloud targets” and directional distance transform (DDT), which is used to remove the expected path of shadows based on sun-sensor geometry.
