Field notes: the pass that caught a slick nobody predicted

We'd been running the same drift model output for three days, watching a convergence line march slowly south along the frontal boundary. HYCOM currents, ECMWF winds, a Lagrangian particle run seeded off last month's known accumulation. Everything pointed to a tongue of aggregation forming around 34 nautical miles south-southwest of where we'd last tasked a vessel. The kind of forecast that makes a cruise plan write itself.

Then a pass came back with a windrow sitting almost sixty miles northeast of that line, in water the model had flagged as low-probability dispersal. No convergence signature in the surface current output. No obvious frontal feature on the SST composite from the night before. Just a long, broken streak with the spectral signature we've learned to trust, sitting in a spot nobody on the team would have picked.

This is the part of the work that doesn't make it into the method section. Drift models are good at telling you where debris should collect given the physics they can see: geostrophic flow, wind stress, Stokes drift if the run includes it. What they're bad at is the transient stuff. A short-lived eddy that spins up for eighteen hours and dies. A local wind shift that sets up Langmuir cells in a patch of water the regional model treats as uniform. None of that shows up until the surface itself tells you.

We've started keeping a rough tally of these mismatches, more out of habit than rigor. Most of them resolve in the model's favor eventually, meaning the pass caught something the forecast would have found a day or two later anyway. But a meaningful fraction don't resolve at all. The slick sits for a tide cycle, breaks up, and is gone before the next cloud-free pass comes around. If we'd waited for the model to converge on that location, we'd have been scoping an empty patch of water.

The instinct, when the model and the imagery disagree, is to trust the model. It's built on physics, it's been validated against drifter buoys, it has a track record. The imagery is a single snapshot, noisy, sometimes confused by sun glint or a thin cirrus deck. But a snapshot of the actual surface beats a forecast of where the surface should be, every time the two disagree about something transient enough to matter for a one-week cruise window.

What changes, practically, is the order of operations. Instead of letting the drift model set the cruise plan and using imagery to confirm it on the way out, the unplanned detection becomes the thing you plan around. You check it against wind history for the prior 48 hours, see if there's a plausible convergence mechanism even if the regional model missed it, and if it holds up, that's where the vessel goes, model line or no model line.

None of this says drift models are wrong to use. They're the only way to anticipate aggregation before anything is visible on the water, and most cruise planning still has to start there. But treating a cloud-free pass as a check on the model, rather than the other way around, has caught more real slicks for us than waiting for the forecast to agree with what's already sitting on the surface. Plastic Detection is built around that ordering: event-driven passes that flag a windrow when the spectral signature shows up, whether or not a drift model called it first.

If your team is still planning cruises off forecast lines alone, it might be worth seeing what the next pass actually catches.