Red tide has turned long stretches of Florida’s coastline into a rotting fish carcass dump, and worse. For the past 10 months, the algae bloom has killed off wildlife, kept tourists away, and caused health problems in locals. The damage grew so widespread this summer that Florida’s governor declared a state of emergency last month. A computer vision algorithm could help limit some of the effects by determining red tide hot spots faster than current methods.
The culprit behind this particular algae bloom is a single-celled dinoflagellate called Karenia brevis, which releases neurotoxins after dying or being broken down by ocean movements. Those neurotoxins can kill animals–it is deadly enough that it’s being blamed for the death of a whale shark–and cause health issues in people. The neurotoxins in K. brevis can also go airborne, causing respiratory problems in beachgoers.
After obtaining a water sample and putting it on a slide, volunteers–all currently retirees who have been trained by Mote Marine Laboratory–take a close-up video, using a $500 setup made up of an iPod, microscope, and a specially designed 3D-printed adaptor that connects the two. The footage is uploaded to the cloud where analysis by a computer vision model detects whether K. brevis cells are in a clip or not. If they are present, HABscope does a count to determine the amount of red tide at that location.
That information, along with other data points like wind conditions and water currents, goes into another model created by NOAA that predicts breathing conditions for the next few days at a particular beach. The sped-up data collection and analysis will allow NOAA to create daily, instead of weekly, alerts, which they hope to roll out later this year. While the warnings don’t do much for sea life, tourism is a huge part of Florida’s economy. Improved forecasts can help people at risk of developing health problems better navigate beach trips.
The biggest challenge was gathering data. There was no data set for K. brevis, so HABscope’s engineer Bob Currier had to collect 50,000 images to feed into a model built using the open-source Tensorflow framework.
“It’s not like training an algorithm on cat photos,” Currier says. “There were zero images out there.”
HABscope could also be trained to spot different algae bloom-causing cells, which have different shapes, vary in sizes, and have unique attributes that computer vision can differentiate. Other communities affected could gather their own data sets, use the technology, and put it to work.
“Moving science forward has to be collaborative efforts,” Kirkpatrick says.