A mask that is too eager throws away good data. One that is too permissive poisons a model. Where we landed, and why.
A cloud mask decides what a customer never sees. That makes it a product decision wearing an engineering costume, and it deserves to be argued about in those terms.
Too eager throws away good data
Our first mask was tuned on precision. It flagged thin cirrus, haze and the bright edges of a harvested field, and it was right often enough to feel good in a notebook. In production it deleted about nine per cent of usable imagery over Northern Europe in winter.
Too permissive poisons a model
The opposite failure is worse and quieter. A model trained on lightly clouded tiles learns that a particular kind of white blur means healthy canopy, and nobody finds out until a season with unusual weather.
Where we landed
Two masks, both shipped. A conservative one is on by default; the aggressive one is a flag. Every tile also carries a per-pixel confidence band, so you can make the call yourself instead of inheriting ours.