Fig. 2

The overall architecture of the method proposed in this study. Step 1 is the Semantic-to-Pixel framework for mapping of multiple water body types. Step 2 is the proposed weakly supervision method. It contains a student network and a teacher network, where the teacher is momentum-updated with the student. The labeled data from the baseline-year is directly fed into the student network for supervised training. Given an cross-year unlabeled image, we first use the teacher model to make a prediction, and then separate the pixels into reliable ones and unreliable ones based on their entropy. The reliable predictions are directly used as the pseudo-labels to advise the student. Besides, the subordinate relationships between the water bodies and their various categories are also used to constrain the network’s training.