The Lab
Where we build, test and sometimes reject. This is the R&D behind VisScanner — open about what works, and just as open about what doesn't.
Our rule is simple — we call it the refutation gate: a model that doesn't hold up on data it has never seen doesn't ship. The same goes for findings. We only publish once we can no longer knock down our own conclusion. This page shows what that process sometimes looks like.
Development
The Rebuilt Forecast
The forecast page has been rebuilt on ground that holds: per species, per day, as a percentage above or below a normal day that month. Wind, moon phase and air pressure are gone. Not because they did not fit, but because they did not earn their place. Plus a new big-fish atlas, language in the URL, and a noticeably faster site.
Research
Our model predicted anglers, not fish. What survived was season and water temperature
For two weeks our model said wind was the biggest driver of whether you catch fish. It was not. It was the shape of the Dutch coastline. Then the day-level signal turned out to know mainly which water is good, not which day. Here is what we found when we audited our own numbers, and how the forecast page was rebuilt around the answer.
Research
Eel in the dark, pikeperch in the moonlight: teaching the forecast to tell time
We're extending the catch forecast with temporal features — hour, season, moon. Most are small on average, but one splits the night hunters: eel wants the dark, pikeperch the moonlight.
Articles
From dataset to waterside: our founder is joining ATKB on a fish survey
Testing models against reality eventually means: boots on the ground. Wytse Punter is soon joining a fish-stock survey run by ATKB.
Research
Teaching a model to see what an angler sees: self-supervised image recognition on aerial photos
Overhanging trees, bridges, T-junctions, lily beds. We're teaching an AI model which structures go with which fish — and why we haven't published it yet.
Research
Reading the bottom from space: why we don't roll out satellite depth everywhere (yet)
We reproduced the standard method for satellite depth and built a model around it that works where the classic approach breaks down on turbid inland water — and still decided not to ship it everywhere. Here's why.