A collaboration of field biologists and computer engineers in Panama aims to create a sound-based counting tool for the endangered manatee, which often lives in opaque, vegetation-filled waters but which researchers hope can be identified individually by its calls.

Among those involved in the effort is Fernando Merchán, an engineering professor at the Technological University of Panama. Merchán has been using algorithms to parse 1,446 vocalizations of the Greater Caribbean manatee (Trichechus manatus manatus), a sub-species of the West Indian, or American manatee (Trichechus manatus). The single-note calls, which were collected in the wetland waters of Panama’s Caribbean province of Bocas del Toro, last roughly one-third of a second and sound like something between a chirp and a squeal.

Merchán has been testing various computer algorithms on his large audio database, the aim being to determine which algorithms most reliably match randomly selected calls with the manatee that uttered them. Ultimately, the goal is to create an affordable computer-aided tool that can more accurately estimate the size of manatee populations in the murky wetland waters where Panama’s manatees are found. Previous methods involving aerial monitoring or sonar have proven ineffective or too expensive.

If successful, the effort could not only boost manatee-conservation efforts, but also mark a step forward in the fast-growing field of machine learning in audio monitoring of wildlife populations, experts say.

Research by Merchán and four colleagues has shown that a machine-learning tool can identify individual manatees through analysis of their vocalizations with over 80% accuracy. The findings were published in August in the journal Frontiers in Marine Science under the title “Unsupervised identification of Greater Caribbean manatees using Scattering Wavelet Transform and Hierarchical Density Clustering from underwater bioacoustics recordings.”

The database of manatee sounds is a product of hundreds of hours of field work by researchers from the Smithsonian Tropical Research Institute (STRI), a Smithsonian Institution center of tropical-ecosystems investigation that is located in Panama. The project involved capturing and recording 23 manatees over a two-year period, says Héctor Guzmán, an STRI staff scientist who led the field team and coauthored the Frontiers in Marine Science study.

Guzmán’s team built a floating four-meter-by-four meter cage using plastic pipes and a fishing net and placed it in Bocas del Toro’s San San River. Manatees were attracted to the cage by a bucket of banana leaves and pulp. Once the animal entered, the team closed off the entrance and attached a hydrophone and digital recorder to continuously record the manatee’s vocalizations for a six-to-10-hour period.

The researchers gathered an average of 63 vocalizations per manatee during these sessions. They also measured the length of each animal, recorded its sex and photographed any scarring or other marks for identification purposes. Given the length of the sessions, Guzmán says, the team spent several days at a time in the wetland environment, often beginning work at three in the afternoon and finishing in the pre-dawn hours of the following day.

“There were tons of mosquitos,” he says. “There was lightning, there was thunder. And that’s the way it works when you’re in the field.”

Guzmán, who has worked with manatees for 14 years, says the manatees use their vocalizations for a wide variety of purposes. They vocalize during courtship, to signal danger, and most of all in communications between mothers and their offspring.

“We have very beautiful records of mothers and calves,” Guzmán said. “It’s a very neat relationship. And they are always chatting, put it that way. They are communicating to each other—blah blah blah—and it’s funny. It’s kind of [as if] the mother is always complaining, and the kid is always complaining.”

Training a computer in Panama City to analyze manatee vocalizations captured in distant wetlands, Merchán has used random samples of the entire database to test whether certain audio-analyzing algorithms can correctly connect each individual chirp with the manatee that made it. The algorithms are examining mathematical representations of the vocalizations based on features such as pitch and length.

“The algorithm works blindly,” Merchán said. “I don’t tell the algorithm these are the correct vocalizations from manatee one, two, three—no. It works blindly. Then I compare the ground truth, the original labels and identification associated with each vocalization.”

In a series of digital experiments, Merchán first extracted random tranches of sound from 10 to 20 different manatees and 10 to 50 of each manatee’s vocalizations. He fed 100 of these randomly generated datasets into different machine learning algorithms in order to determine which ones were most successful in matching calls to the individual manatees that uttered them. The best-performing algorithms underwent further testing with another 100 sample datasets, this time expanding the range of individual vocalizations to 30 to 50. The most successful algorithms were once again tested against the entire data set of 1,446 vocalizations, with the best-performing machine learning technique achieving 83.75% accuracy.

The recently published study builds on past collaboration between Merchán and Guzmán. The two had previously used recordings from stationary listening devices in these wetlands and machine learning tools to estimate manatee populations. (See "In Panama, manatee researchers listen in" —EcoAméricas, July 2020.) This more recent study, however, allowed them to test the algorithms’ effectiveness in distinguishing among precisely identified manatees.

Merchán and Guzmán are now hoping a more finely tuned analysis will help them estimate populations using 10 years of stationary manatee recordings that they collected at Bocas del Toro, which is located in northwest Panama, and in the Panama Canal.

“That’s the purpose of all this,” Guzmán says. “We’re not like in Florida, where you jump in the water, you can see the animal and even touch the animal, et cetera. That’s not our case unfortunately.”

The Panama research builds on work stretching back decades, including investigations by Brazilian biologist Renata Sousa-Lima. She authored studies published in 2002 and 2008 on the Amazonian manatee (Trichechus inunguis) and the Greater Caribbean manatee, respectively. Both manatee species communicate with similar chirp-like calls.

Her 2002 study showed that Amazonian manatees can identify each other based on their calls. Drawing on further research, Sousa-Lima authored a 2008 study that showed differences in vocalization pitch, length and other features based on the manatee’s gender, age and size.

“The first thing that the calves do when they’re born is to call for mom,” she says. “And their mom responds and then they learn each other’s voices.”

Sousa-Lima notes that this distinguishing continues later in life, for example when adult males seek out females. She echoes Guzmán’s observations about the difficulty of tracking the Greater Caribbean manatee due to it favoring dark, vegetation rich waters, where visual sightings are difficult.

The Amazonian manatee is hard to count visually, too, she says. The subspecies in Brazil also favors vegetation-rich waters, which in the Amazon region tend to be opaque due to silt and tannins. The fact that the Amazonian manatees don’t migrate adds to the difficulty, she says, since their population remains spread out. In Florida, by contrast, more pronounced seasonal changes prompt manatees to migrate to specific locations, where they aggregate and can be counted relatively easily, she notes.

Sousa-Lima continues sound-based work as a professor of animal behavior at the Federal University of Rio Grande do Norte, based in the city of Natal. There she founded the Laboratory of Bioacoustics (LaB) in 2011. LaB’s recent projects include research similar to Guzmán’s—combining audio recording and machine learning analysis across a wide range of wildlife including bees, birds, fish, whales, dolphins, bats, ants, and turtles.

“Turtles actually speak when they’re inside their eggs,” Sousa-Lima said of one study her laboratory helped conduct. “And they coordinate among themselves—’one-two-three, let’s hatch.’ And when they do hatch and go to the water—[in the case of] the Amazonian turtle species—the moms kind of sort of wait for the babies to come in so they can migrate together. And that’s all mediated by sound. And so, there’s much more to aquatic marine communication than we knew before.”

She says dramatic improvements in audio-recording technology have spurred wildlife scientists’ use of machine-learning tools. The move was a natural step since machine learning already was being used to identify individual animals from visual footage, she notes. For field researchers, collecting years of audio recordings, as Guzmán and his colleagues did in Panama, is becoming commonplace.

“Now we have big, huge, enormous, gigantic data to deal with,” Sousa-Lima says. “And we can’t listen to it all, you know, manually.”

Of the two subspecies of the American manatee, the Florida manatee (Trichechus manatus latirostris) has a relatively small range that is limited largely to the southeastern United States. The range of the Greater Caribbean manatee is far larger, stretching from the eastern coast of Mexico, through Central America, and across the Caribbean coast of South America, all the way to the easternmost regions of Brazil. Caribbean islands including Cuba, Puerto Rico and Hispaniola are also home to this subspecies.

In 2008, the International Union for Conservation of Nature (IUCN) listed the American manatee as vulnerable, citing an expected decline of 10% of its population over a 60-year period. The lifecycle of the manatee contributes to its vulnerability. Mothers typically give birth to a single calf at a time, after a gestation period of approximately 11 months. Then they nurse their calves for approximately two years, which means depleted populations take considerable time to rebound.

The Greater Caribbean manatee subspecies was estimated by IUCN at only 2,500 adult individuals in 2008 with an expected decline of more than 20% over the next 40 years. Habitat loss, hunting, and accidental deaths due to boat strikes are cited as the principal threats.

One of the larger hurdles facing conservationists is that reliable population estimates are difficult to obtain, an issue that audio-based research aims to resolve. Visual counts based on aerial surveys in Puerto Rico (2019) and Brazil (2015) ended up with wide estimated-population ranges. A pair of 2017 studies in Mexico also note the difficulty of counting manatees using current methods. Not coincidentally, Mexico does not have a nationwide manatee-population estimate.
Panama granted the manatee protected status in 1967, and in 2023 passed a law creating protections for manatees and other aquatic mammals in a “marine corridor” that hugs the Caribbean coast. Lethal captures were outlawed, and non-lethal captures can only occur for scientific or medical reasons approved by the country’s Environment Ministry.

In Brazil, the Greater Caribbean manatee is classified as critically endangered within the country, according to a 2022 study in the Journal for Nature Conservation. The authors argued for this highest-risk classification based on past and current hunting threats, boat collisions, habitat degradation, and incidental deaths from fishing.