The tropical ecosystems of Costa Rica and Puerto Rico have ears, and have done for some time. These recording stations were put together with iPods and car batteries which each record 144 60-second recordings every day, and transmit them to a web-enabled base station up to 40 km (25 miles) away. From there they're uploaded to a web app with which biologists train a software algorithm to recognize the chirrups, squeaks and caterwauls of the forest's birds, monkeys, frogs and other fauna. It's all in the name of documenting wildlife, to better understand the effects of deforestation and climate change. And according to scientists at the University of Puerto Rico, it sure beats putting boots on the ground.
The problem with conventional monitoring, Dr. Mitchell Aide explains, is that field experts are limited in number. They cannot be everywhere all the time, and so the data available is necessarily limited. "To understand the impacts of deforestation and climate change, we need reliable long-term data on the fauna from around the world," he says.
To do that the team devised an automated remote recording station built with off the shelf parts. A microphone connected to an iPod via a pre-amp is the actual recording device, powered by a 12-V car battery and solar panel. Most of the electronics are protected from the weather by a sealed case, with a 900-MHz antenna attached to transmit one 1-minute recording every 10 minutes to the nearest base station. With a clear line of sight, recording stations can transmit up to 40 km, but in thick vegetation, this is slashed to 2 km (1.2 miles).
The team says that with the intermediate gain settings used, the recording station is able to detect the common coqui, a frog native to Puerto Rico from a range of 50 m (165 ft), giving the station a 1-ha sampling area for that particular species.
The base stations, which receive the recordings before losslessly compressing them for upload to the web, are essentially Mac minis with an external hard disk connected to another antenna and a network switch.
From the base station, the recordings are uploaded to the web where they can identified, logged and organized. But more than that, the scientists hope that users of the system will help to train it to identify particular species on its own, the better to monitor the longterm fortunes of wildlife. Though anyone is free to visit the website to listen to recordings minutes after they were captured, only approved users may analyze them.
Though this team's work is focused on Costa Rica and Puerto Rico, the website lends itself to any project based on accrued audio recordings (the logged recording count from these countries as well as the United States, Brazil and Argentina is over 1.3 million.)
Using the web interface, recordings are annotated by the user. This is done for every call of every species, so that the system will learn (using the Baum-Welch algorithm) to identify that particular call, and find and log matching examples. Correct matches are incorporated into the training data, and incorrect ones discarded, until the user is satisfied that they've developed an effective model. The whole system, from field recording equipment to web app, is called the Automated Remote Biodiversity Monitoring Network, or ARBIMON.
The team demonstrated ARBIMON with recordings for nine different species from a range of taxa based on five years of recording data from its original recording station in Sabana Seca, Puerto Rico. Impressively the system took under 2 hours to run three models for the 170,000-odd recordings. The team reports that the accuracy of the system ranges from 79 to 99 percent depending on the species (highest for the strawberry poison-dart frog species Oophaga pumilio and lowest for a particular but unknown insect species).
Though accuracy, i.e. the correct identification of the species, was generally high, the team reports that precision (or the hit rate for identifying the species' presence at all) was more of an issue. False negatives proved more of a problem than false positives, with only one instance of Great Tinamou (Tinamus major) calls identified by the model out of the 41 recordings available to it. "These results suggest that these models are relatively conservative; they rarely confused the species with another, but they do not always detect the species when it is present," the team finds in its paper Real-time bioacoustics monitoring and automated species identification published on Tuesday in the online journal PeerJ.
The team was also able to pinpoint the times of day the species monitored were most active, as well as fluctuations in activity month on month and year on year. For example, in the case of the recently-identified endangered frog species the plains coqui (Eleutherodactylus juanariveroi), the team identified a four-year decline in activity followed by a fifth year recovery, "a dynamic which is common in many species, but often difficult to measure," it writes.
Far from making biologists redundant, the researchers hope the ARBIMON system will improve their lot. "We are trying to provide the best data and tools possible, so that the biologists can use their time to convert these data into useful information for science, conservation, management, and education," says Dr. Carlos Corrada-Bravo. And because the data is a permanent record, future scientists will be able to apply the analytical methods of tomorrow.
"Conserving and managing the biodiversity in the world is a major challenge for society, particularly in the tropics," adds Dr. Aide. "We hope that the tools we have developed will allow researchers, students, managers, and the public to better understand how these threats are impacting species, so that we can make informed conservation and management decisions."
If ARBIMON proves its worth, denizens of the world's most remote ecosystems may increasingly find themselves the unwitting targets of electronic ears, constantly drip-feeding their noises to a machine mind, itself quietly going about the business of tracking the world's wildlife.
See the stories that matter in your inbox every morning