PAM Glider Rodeo Hackweek

Noise Nowcaster

Predicting how loud the ocean is — and how far a glider can hear because of it.

Suramya Angdembay · Kapil Sharma · Bibas Kandel
with the AQUAVIEW team

Built on AQUAVIEW & Glider Rodeo data

5 minutes

A noisy hour and a quiet hour are not the same survey

Two circles on a dark background. The quiet-hour circle is twenty-three times the area of the loud-hour circle.

A glider hears whales only as far as the background lets it. Over one mission the sperm whale band swung about 14 dB — and the water it was effectively covering moved with it.

Measured noise over the mission; area compared under a simple spherical-spreading approximation

Where the data came from

Two sources, one table

Map of six glider tracks over bathymetry contours off the west of Oahu, with Kaena Point marked and a ten kilometre scale bar. All six tracks overlap in the same small area.

AQUAVIEW / PacIOOS

every environmental input — about 80% of what the model sees

Rodeo hydrophones

the measured noise levels — the other 20%, and every label

6 gliders and drifters

1,565 experiment-qualified hours

5 whale-call bands

The test we held ourselves to

Every score is on a glider the model has never seen

Train on five platforms, score on the sixth, rotate. Every number in this talk is from a glider the model had never met.

Two time-series panels comparing measured and predicted noise for two gliders the model never trained on. The prediction follows the broad rises and falls; one panel tracks closely, the other less so.

Why we split it

A calm sea and a rough sea are two different problems

Scatter of noise against wind speed. Below about 6.5 metres per second the points are widely scattered and the fitted line is nearly flat; above it the points tighten and the line is about two and a half times steeper.

Wind is the strongest driver in the whole model — and it found that on its own, nobody told it. But it drives the noise differently depending on how hard it is already blowing.

Our method — a follow-up experiment, not the shipped tool

So we trained two specialists and a switch

Diagram: this hour's wind, waves, swell and currents feed a switch that asks how rough the sea is, which routes to either a calm-sea expert or a rough-sea expert, and both feed the predicted noise for five bands.

Neither specialist has to be good at everything — which is the whole point.

A follow-up experiment, not the shipped tool

Two specialists and a switch — but the switch has to look at the right thing

Bar chart: a switch based on sea state improved predictions by 2.8 percent, while switches based on extra physics, location or bathymetry produced no real change.

One model learns calm conditions, the other learns rough ones.

For each hour, the switch decides which to trust.

Switching on where the glider was did not improve things. Switching on what the sea was doing did.

Against our own first model

Better on a glider it had never seen

Bar chart of prediction error on an unseen glider. The baseline is highest. Fine-tuning is 2.3 percent better, sea-state experts 2.8 percent, per-platform rescaling 3.7 percent and both combined 4.4 percent, but only the sea-state experts held up when re-run.

Baseline here is the same model without the switch — so the only thing that changed is the routing. Against the team's original model it is 4.7%, and against assuming normal conditions, 13%.

The tool you can run today uses the gradient-boosted model. The experts model is a follow-up result, not yet wired into it.

It helps when meeting a new glider. It does not help forecasting further ahead — there it is worse.

Where it works, and where it doesn't

Sea state carries real signal higher up — and almost none at 20 Hz

Grouped bar chart by whale-call band, drawn against a 100 percent ceiling marked predicting the hydrophone exactly. The sea-state expert model explains 11 percent of the swing in the fin whale band, 27 percent for humpback song, 34 percent for odontocete whistles and 33 percent for beaked whale upsweeps, each above the baseline and all far below the ceiling.

The fin whale band sits at 10–30 Hz, where vessel noise matters most — and vessel traffic is the one major input we did not have. We have not isolated shipping as the cause, but it is the clearest thing to try next.

Conclusions

Score on a platform you have never seen. Anything easier will flatter you.

Sea state is what separates the regimes. Location and depth are not.

Get vessel traffic data. It is the missing piece for the low-frequency bands.

The simple fix usually matched the clever one. We shipped the simple one.

Code, method and caveats are in the repo

Thank you