About Bim Beach

Bim Beach publishes two experimental products for the waters around Barbados: an offshore Sargassum drift forecast, and Sandpiper — Beach Watch, a record of how much wrack is sitting on surveyed beaches.

Both are research outputs made public. They are not an official service.

Sargassum tracker

Sentinel-3 OLCI sees Sargassum rafts offshore roughly once a day, when cloud allows. The tracker seeds particles on those detections and advects them forward for 72 hours through a high-resolution coastal ocean model (FVCOM) nested inside the US Navy's AmSeas regional model, with fitted windage and Stokes-drift terms.

Particles are not re-confirmed once the forecast starts, because a forecast has no future satellite pass. Where cloud hid the sea at issue time, particles carried forward from earlier images fill the gap — the model is then the only source of information there, and the page says so.

Reading the tracker map

The loop is 144 frames, one every 30 minutes of forecast. Drag or scroll to pan and zoom; the scale bar and range rings follow the zoom, so distance stays readable from the whole 260 km view down to a few kilometres.

The clock at the top left is the forecast age. Nothing on the map is re-confirmed by satellite after it starts, so a particle at +72 h has been carried by the model alone for three days.

What the tracker is for

A sense of direction and timing offshore — whether material is heading toward the island or away from it, and roughly when. It is not a landfall forecast.

Sandpiper — Beach Watch

Beach Watch estimates how much of each surveyed beach is covered by wrack, from Sentinel-2 imagery, across nine beaches. Each beach carries its own measurement error, and some beaches are much noisier than others.

Crews clear wrack from the south and south-east beaches, so cover measured there is arrival minus removal, not arrival.

What a Beach Watch card means

The big number is the percent of the scored window covered by stranded wrack. The headline word (clear / light / moderate / heavy / severe) is the band that percentage falls in, so the colour and the word can never disagree. The trendline is every cloud-free look on record, positioned by date, so a fortnight of cloud reads as a gap rather than being closed up.

Hovering a card shows the scored window. That is deliberately not the raw map polygon: the outline is rasterised onto a fixed box, so any part of it beyond the box is never measured — for Atlantic Park only 24 of its 305 outline points fall inside its own box. Permanent vegetation, rock, cliff and water are masked out and never scored, and the popup states how much of the box that removes.

Accuracy, stated plainly

Every beach carries its own error, measured by fitting the model without that beach and then testing it on it. Across 32 labelled scenes the model is off by 9.2 percentage points on average, against 13.4 pp for simply predicting the average of every other beach. So ranking beaches is far better supported than any single percentage. Cards whose own error exceeds ±15 pp are hatched on the page, and the per-beach error is a column in the summary table.

It discriminates. Held out by beach, the detector reaches AUC 0.79 for whether there is any wrack at all and 0.81 for whether it is moderate or worse, and beats a constant on Brier score in every combination. Rank correlation against the human labels is +0.62.

It is not calibrated, and cannot yet be shown to be. Scored with ten equal-width bins on 32 scenes, a perfectly calibrated predictor produces an expected calibration error of 0.16 by binning noise alone. Every value measured here sits inside that. The test has no power at this sample size, and reading it as "well calibrated" would be measuring the estimator rather than the model.

An offseason baseline was tried and failed. Subtracting each beach from its own clean 2018–19 median was meant to remove permanent beach character. It tripled a beach labelled zero four times, and left Atlantic Park's winter level untouched — a dry-season baseline sits at a different sun angle from a summer scene, so the subtraction adds a season-shaped error in phase with the season it is testing. What survived is the static mask built from the same baseline, which cut the held-out error from 10.8 to 9.2 pp.

The east-coast dry season is unexplained. Atlantic Park and The Soup Bowl read 17–19 % in winter, when Sargassum is at its minimum. That survived both the mask and the subtraction. It is either real winter drift on the Atlantic shore or a beach property neither test reached, and it is stated here rather than smoothed away.

Carlisle is the control. Every one of its labels is zero, across both label samples. It must stay flat, or the whole record is measuring something other than wrack.

What this does not do

This section is the important one, and it is deliberately blunt.

• It does not predict which beach Sargassum will land on. Testing against an independent record of 791 measured beach-cover values returned an indeterminate result: transport was not shown to beat either a wrong-date control or a simple "how much Sargassum is offshore today" baseline.
• It does not resolve the nearshore. The model's coastline sits 2–5 km offshore. A particle reaching it has arrived at the kilometre-scale coast, which says nothing about which beach, or whether anything comes ashore at all.
• It does not estimate tonnage, mass or concentration. Brightness on the map encodes how many model particles stacked, not how much weed is there.
• It is not validated as a forecast. Position error grows with lead time and has not been established beyond short spans.

How often each product updates

Sargassum Watch. Forcing data is captured daily. A forecast is issued when the inputs and the model run both succeed; when they do not, the page keeps showing the last good result and marks it stale.

Beach Watch. Updates when Sentinel-2 publishes a usable scene — not daily. The satellite revisits about every five days and cloud removes most of those passes, so a card records when that beach was last seen, not what it looks like today. A card several weeks old is normal rather than a fault; the page says how old it is and marks it stale past a fortnight.

Contact and corrections

Errors and corrections are welcome. This is a research product and treating it as one is the appropriate posture.