Data quality control procedure

Data quality control is an important step before public release of data. Most quality issues occur because of errors generating either the monthly composites or the annual median, often due to false negatives in the cloud detection process or artifacts in the Sentinel imagery themselves, such as image saturation near the edges of orbits.

While we prefer to keep all the predictions to be in the year of interest, between 0 and 2% of the data may rely on 2019 data.

1. The first pass is to run the 2020 model with algorithmically determined (temporal or median) predictions
2. Then, select tiles with quality control concerns and rerun predictions with image resegmentation pipeline
3. Select tiles with quality control concerns remaining after (2), and rerun predictions with 2019 model
4. Select subtiles with quality control concerns, and set subtiles with quality control concerns to no data flag (255)