iCorrected viewColour correction runs in the cloud once the image arrives, not on the
camera. An early test — results will improve.
…—
Corrected——
Corrected
—
iFish count — model estimates
Each frame was sent once to a vision-language model, which returned a box and a confidence for every fish
it saw, and a total count. Show inference draws those boxes with their confidence and names the
model that produced them. Results are pre-computed; this page never calls the model.
Confidence is self-reported by the model, not a calibrated probability. Boxes are
approximate. Proof of concept.
Min confidence0.00
visibilityHow this is derivedEach capture is converted to greyscale and passed through an edge filter; the
score is the mean edge response across the frame — mean(|∇I|). Suspended sediment
and plankton scatter light, softening every edge, so the number falls as the water thickens.
No model. Observed 0.8–17.5. A proxy, not metres. A partial frame is missing
pixels, which lowers the score without the water being worse — the % beside it is completeness.clearhazymurkycamera offcamera silent (no capture for 30 h)deployment
Conditions & history
◆ deployment● notefish countwavesSignificant wave heightsignificant_wave_height_m from the Sofar Spotter on this
mooring, hourly, around the clock.The buoy reports through the night while the camera
sleeps, so this series is continuous where the image series is not.windWind speedwind_speed_mps, same Spotter, hourly.
Across this deployment wind and visibility show no meaningful relationship
(r = +0.14). Context, not a predictor.sea tempSea surface temperaturebm_soft_temperature_mean_13bits from the Sofar smart-mooring
temperature sensor on this buoy: one reading an hour, at :10 UTC.pHpH — SAMI dev kit (provisional)A Sunburst iSAMI-pH on Bristlemouth node 0xc26a…2d6a
measures once an hour. Its 466-character record reaches the buoy as three messages; the backend
reassembles them and converts to pH with the public bm_iSAMI firmware's algorithm (salinity 28), which
matched NOAA's own reading for 7 Oct (7.98). NOAA's build and coefficients are not
confirmed, so the absolute value is provisional — compare with coral.aoml.noaa.gov/mir.
Hover the readout for the raw values and the textbook variant (salinity 36, pKa at the in-water
thermistor). ▲ marks an hour where one of the three messages was lost.missingMissing dataHatched where a sensor skipped a reading: more than 1.5× its cadence between two
samples, or since its last one. Captures count as silent after 30 h without one — a missed duty day.
A gap on every lane at once is the buoy or its uplink; a gap on one lane is that sensor
or its cable.
This is a demonstration site showing a proof of concept, not a finished product.
Imagery is published with the express permission of our deployment partners.
Live data. Captures and completeness come from /devices/<id>/media; waves, wind, sea
temperature and pH from /systems/<id>/spotter-telemetry (streams waves, wind,
soft_temp, sami_ph; sami_ph is served to the admin token only, so the public demo
never shows it). All times are stored UTC and rendered in the selected zone.
Visibility could not be measured — the image host does not allow pixel access
from the browser, so capture bars are coloured by transmission completeness instead. Computing it at ingest
would make it exact and free here.
Events and deployments are stored server-side and shared by everyone with access to this system.