- Developing effective management strategies for threatened species like koalas requires knowing where and how many are in a target area, but surveying cryptic low-density animals can lead to variable estimates.
- A recent study has introduced a new automated method for wildlife detection using a pair of object detection machine learning algorithms to detect animals’ heat signatures in drone-derived thermal imaging.
- By understanding error rates of different survey methods and including appropriate technology, the researchers say, wildlife monitoring can become more efficient and effective.
Koalas are Australian icons, known globally for their fuzzy ears, broad wet nose, and love of eucalyptus leaves. They are also cryptic, slow-moving creatures that live high in the leafy canopy of said eucalyptus trees and are threatened by human activity.
Developing effective conservation management strategies for threatened species like koalas (Phascolarctus cinereus) requires knowing where they are and how many are in a target area, so scientists need be able to accurately survey and monitor their populations.
Accurate monitoring via surveys requires not only finding the animals but knowing how easily and consistently they can be found with the given method; for statistical analysis, it means being able to estimate the probability of detecting them on any given survey.
A team of researchers from the Queensland University of Technology (QUT) in Australia has done just that for a new method combining a trio of increasingly popular wildlife survey tools: small drones, thermal imagery, and machine learning. The researchers found that their combination method more accurately detected the cryptic koalas than expert observers did, while saving time and money.
New tools for wildlife surveys
As small, commercially available drones become cheaper and more reliable, scientists are exploring their potential as a tool for a variety of wildlife research and monitoring activities. Drones offer a unique aerial viewpoint, and their ability to carry cameras or other sensors as they move overhead enable them to scan an area from a range of angles.
Thermal imagery captures heat signatures, making the warm bodies of animals stand out against cooler vegetation, air or soil. That makes thermal imaging ideal for surveys of cryptic, hidden and nocturnal animals.
The QUT researchers evaluated the ability of machine learning algorithms to detect koalas in thermal images of the animals’ natural surroundings collected using drones. In their recent paper in the Nature journal Scientific Reports, the researchers write that determining the accuracy of automating wildlife detection methods, particularly for cryptic species living at low densities in complex environments such as woodlands and forests, “remains an open challenge” that their new technique addresses.
“The really exciting thing scientifically is that we have an accurate monitoring method for an animal that lives in a complex habitat,” Grant Hamilton, a senior lecturer at QUT’s School of Earth, Environmental and Biological Sciences, told Mongabay.
Several prior studies have tested the ability of automated analyses to detect target animals in drone-based imagery taken at specific speeds and heights. These include counting the number of seabirds on a beach which machine learning analysis of drone-based images did more accurately than expert ground observers or human image analysts could; measuring the body size and mass of leopard seals resting on Antarctic ice; detecting large-bodied shallow-water marine species just below the water’s surface; and surveying European nightjars in logging or utility concessions the U.K.
Preparing field work for automated analysis
Koalas’ quiet, cryptic nature and the complex structure of their eucalyptus homes makes them difficult to find from the ground. The new study builds on earlier related work by QUT researchers that served as a proof-of-concept of koala detection using drones and thermal imagery; equally importantly, it also determined that drone flight patterns affected detection algorithms.
“The speed and height at which you fly and the sensor you use interact to feed into a machine learning algorithm,” Hamilton said, “so the actual pattern in which you fly is relevant goes into the algorithm.”
That initial project could detect individual koalas from above, Hamilton added, but it didn’t focus on developing a complete methodology that included not only the drone but the capacity of the machine learning to consistently and accurately detect the koalas in the thermal images.
“The paper we just published is really about communicating the methodology,” Hamilton said. “It’s about how do you develop a methodology that you can package up and translate out to the world to make sure you get better outcomes.”
In the new study, the researchers assessed the accuracy of this automated approach by counting a larger, statistically valid number of koalas and calculating error rates (how often a person or an algorithm identified a koala that really was a koala). The team used the error rates to compare the capacity of the automated method of counting koalas in the wild to that of traditional ground surveys.
“On average, an expert koala spotter is going to get about 70 percent of koalas in a particular area,” Hamilton said in a statement. “We, on average, get around 86 percent. That’s a substantial increase in accuracy that we need to help protect threatened species.”
The researchers trained the algorithms on images taken from two field sites in Petrie, Queensland, where research teams had previously radio-tracked a population of 48 koalas. The researchers flew the drones in the early morning at 60 meters (197 feet) above the ground, 30 meters (98 feet) above the tree canopy, conducting a total of 11 aerial surveys. The field teams tracked the animals on the same days the drone flights were undertaken, and they recorded their locations with GPS units. This way, the researchers could validate how well the analysis of the drone-based images correctly detected koalas by comparing the pixels in the images deemed “koala” to the actual locations of the GPS-tagged koalas.
To analyze the images, the researchers combined two deep convolutional neural network (DCNN) object-detection algorithms, Faster-RCNN and YOLO. These separately identified objects in the images, the type of each object, and the confidence that the algorithm had in its detection being correct.
The researchers combined the results to identify regions, or thermal signatures, that were likely to be koalas, as opposed to people or kangaroos. If the locations of these heat signatures matched those of the known radio-tagged koalas, they were identified as being a koala. If they didn’t match, they could have been other, untagged koalas, or other similar-sized animals such as large birds or possums.
The process tracked the position of detected koalas over multiple image frames to reduce repeated detections of the same animal and to enable the researchers to manually review each set of detections to see where and when the koala came into view and disappeared from view, comparing these to the GPS coordinates. The researchers recorded the time it took to identify all possible candidate signatures automatically and manually for each survey to determine which was more efficient.
Technology for better, more efficient outcomes
The study showed that surveying the animals with the drones was faster and cheaper than covering the same area with human observers, and was, the researchers wrote, a more reliable and less invasive way to detect koalas. Moreover, combining the two object-detection algorithms to analyze the drone-based thermal imagery provided an accurate way to detect koalas in their complex environments.
“We’ve made a broad comparison of how our method compares to what expert observers can do,” Hamilton said. “We’ve compared a human looking at the thermal imagery and the machine learning algorithm automatically classifying koalas. The machine learning did better.”
The new method detected between 78 and 100 percent of the target species — a higher probability of detection, the researchers wrote in their paper, than that realized in several previous studies that used automated methods to detect animals in the wild from aerial images. These included studies of dugongs, which live at low densities, and bears, which inhabit complex environments, as well as several species that occur in more open environments, such as nesting seabirds and large African mammals.
The researchers also wrote that their automated detection method yielded a higher probability of detection and a lower or comparable rate of false positive error (flagging a koala in an image where there isn’t one) in the same amount of time as manual assessment of the imagery. Moreover, much of the time needed for the automated detection method was computer processing time that could run while researchers did something else.
Measuring and comparing error rates of the automated versus manual detection methods was a key aim of the new study.
Every method has some error associated with it. “What you need to do in ecology,” Hamilton said, “is to understand what that error is for the different methods, and then you can start to understand how your management is having an impact.”
If, for instance, a research team surveys a species several times and finds a widely different number each time, its estimation of the number of animals there could be close to or far from the actual number. If the team repeats the survey a year later and again has high variation in its counts (i.e. large error bounds), they don’t actually know if the population has gone up or gone down because there’s so much error.
“To monitor well,” Hamilton said, “you need to establish what the error bounds are and you need to get accurate methods, and that’s what we’ve done.”
The researchers said that others should be able to apply their methods to survey a range of other species, including those that, like koalas, are sparsely distributed in structurally complex environments and over relatively large areas. Hamilton said the research team planned to apply the new method to detect non-native feral deer, which he said are a problem up and down Australia’s east coast.
Corcoran, E., Denman, S., Hanger, J., Wilson, B., & Hamilton, G. (2019). Automated detection of koalas using low-level aerial surveillance and machine learning. Scientific reports, 9(1), 3208.
FEEDBACK: Use this form to send a message to the editor of this post. If you want to post a public comment, you can do that at the bottom of the page.