- A new computer model developed by engineers at the Florida Atlantic University uses deep learning to count manatees in images captured by cameras.
- The model has been trained to identify manatees in shallow waters, and can be used to identify where they aggregate, which can, in turn, be helpful to plan conservation actions and design rules for boaters and divers.
- However, the model can’t yet distinguish between adults and calves, or between males and females, both of which are details that are vital for conservation and research purposes.
- The engineering team says it plans to continue training the model in the months ahead, while also working with biologists to get their feedback on how to improve it further.
How do you count manatees? Ideally, standing by a river while playing with the aquatic mammals. However, in a world where manatee populations face increasing threats, a faster and more accurate method is imperative.
Enter artificial intelligence.
A model developed by scientists at Florida Atlantic University uses a deep-learning-based method to count manatees in images captured by cameras. A study published in the journal Scientific Reports describes how the model can use even low-quality images to estimate manatee populations in shallow waters.
“Eventually, we think this model will be very helpful to understand manatee demographics in real time,” study lead author Xingquan Zhu, a professor at FAU’s Department of Electrical Engineering and Computer Science, told Mongabay in a video interview. “How many are there? What are their habitats? Where do they go for food?”
Answering these questions is helpful to plan conservation actions, prevent habitat loss, and design rules for boaters and divers.
Manatees (Trichechus spp.) have long been tracked with the help of GPS devices attached to their tails. However, the harsh marine environments they often inhabit lead to the tags breaking down quickly. The use of physical tags also imposes a restriction on the number of animals that can be monitored in one go. “Plus it’s also quite labor-intensive,” Zhu said.
These gaps prompted Zhu and his team to start thinking about adopting artificial intelligence to estimate manatee populations in an affordable and real-time manner.
The first step entailed gathering images with which to train the model. Initially, the team thought of using images from the internet. However, realizing that these wouldn’t be nearly enough, they turned to another source.
“We realized there are quite a lot of videos from the U.S. state parks, and those videos gave us a lot of information,” Zhu said. “They captured manatees on different days, and from different angles in the same spot. They also showed us manatees in different seasons.”
Capturing images from videos was, in itself, a task. To start with, the team had access to thousands of videos, from which they picked close to 700 different frames to train the model. They also had to ensure a good mix of images that showed manatee groups of varying numbers. “In some of the areas, you see only one or two manatees, and in some of them you see close to even 40 or 50,” Zhu said. “We wanted to mix all these types of images.”
With the data in place, the team worked on labeling the manatees in the images. “Labeling essentially means to teach the model, ‘Hey, this is a manatee. This is what it looks like,’” Zhu said.
The model, however, doesn’t directly recognize or understand shapes per se. When a user inputs an image into the model, it takes small patches of pixels from the image and checks it against the labels it was trained on. The model’s algorithm then “basically says ‘Yes, you are correct. No, you are wrong,’” Zhu said. “This process goes on continuously back and forth, back and forth, until the whole image is analyzed.” Once done, the model outputs an estimate of the number of manatees in the image.
However, it’s not always smooth sailing.
Often, the model confuses large fish for manatees. It can also be bamboozled by inanimate objects. “If you put a toy that resembles a manatee in water, the model will recognize it as a manatee,” Zhu said.
Additionally, the model can’t yet distinguish between adults and calves, or between males and females — distinctions that often play a vital role for conservation as well as research purposes.
Zhu said he hopes to work with biologists to improve the model. “The reason we haven’t done that yet is because we were trying to use the computation to see if we got results,” he said. “Now that the results are promising, the next step is to work with biologists and get their feedback.”
He said the team also plans to keep training the model so that it gets better and more efficient. “With deep learning, you can just keep training it,” he said. “There’s normally no end.”
Banner image: Manatees (Trichechus spp.) have long been tracked with the help of GPS devices attached to their tails. Image via Pixabay (Public domain).
Abhishyant Kidangoor is a staff writer at Mongabay. Find him on 𝕏 @AbhishyantPK.
Coordinated community response saves manatees stranded in Colombia
Citation:
Wang, Z., Pang, Y., Ulus, C., & Zhu, X. (2023). Counting manatee aggregations using deep neural networks and Anisotropic Gaussian Kernel. Scientific Reports, 13(1). doi:10.1038/s41598-023-45507-3
Credits
