Fig. 4
From: Dynamic World, Near real-time global 10 m land use land cover mapping
Training protocol used to recover the labeling model. The bottom row shows the progression from a normalized Sentinel-2 L1C image, to class probabilities, to synthesized Sentinel-2. The dashed red and blue arrows show how the labeling model is optimized with respect to both the class probability and synthesis pathway, and the synthesis model is optimized only with respect to the synthesized imagery. The example image is retrieved from Earth Engine using ee.Image(‘GOOGLE/DYNAMICWORLD/V1/20190517T083601_20190517T083604_T37UET’).
