Researchers at Queensland University of Technology (QUT) in Australia are now getting a bird’s-eye view of koalas in their natural environments, thanks to detection algorithms and drones. A study led by Grant Hamilton, a professor in QUT’s School of Earth, Environmental and Biological Sciences, explored an accurate and precise methodology for effective wildlife management that can find koalas in complex structural environments. Using object detection algorithms, the heat signatures of koalas can be detected by drones equipped with thermal imaging. Remotely piloted aircraft systems (RPAS) collect image data of wildlife populations, including koalas. Courtesy of Queensland University of Technology. A large temperature gradient exists between mammals and their environment, allowing computer vision programs to easily pinpoint and count their thermal signatures, according to the researchers. The drones — remotely piloted aircraft systems (RPASs) — were flown over two isolated survey areas in a “lawn mower” pattern. Koala populations are unevenly and vastly distributed within complex eucalyptus canopy coverage, which can make tracking and direct observation of individual koalas less probable. By nature, the behavior of these animals is cryptic, and they inhabit dense environments that reduce the ability of both ground and color photographic imaging to detect them accurately. QUT researchers are better at detecting and studying koalas in their natural environments. Courtesy of Queensland University of Technology. “Nobody else has really managed to get good results anywhere in the world in a habitat this complex and in these kinds of numbers,” Hamilton said, noting that other drone-based animal population detection systems have been used when looking for seals on a beach — a task that is much simpler. “A seal on a beach is a very different thing to a koala in a tree. The complexity is part of the science here, which is really exciting. This is not just somebody counting animals with a drone. We’ve managed to do it in a very complex environment.” The areas were selected based on a population of 48 koalas that had previously undergone extensive tracking and radio-collaring by field ecologists. North and south Petrie in Queensland, Australia, are both bounded by urban areas, high-traffic roads, and a river that inhibits the migration of the populace of koalas. An aerial image of the dense canopy of a koala environmental habitat, obtained with an RPAS. Courtesy of Queensland University of Technology. Eleven RPAS surveys were carried out between February and August 2018, using the Matrice 600 Pro and A3 flight controller equipped with a FLIR Tau 2 640 thermal camera mounted to the underside and a gyro-stabilized gimbal to reduce vibrations in the thermal imaging frames. The drones flew over the sites at a speed of 8 m/s and an approximate altitude of 60 m above ground or 30 m above the tree canopy. All 11 surveys were conducted in relatively the same weather conditions — between 8 and 21.7 °C — and at similar times of day, the researchers said, to provide consistent variables within the data sets. During the missions, ground teams were deployed to validate the data sets from the RPAS-derived thermal imaging. The images were then applied to two published and well-established deep convolutional neural network (DCNN) object detection methods: Faster R-CNN and YOLO (you only look once). Each DCNN was given an input image that produced a list of bounding boxes denoting a confidence level between zero and one (with “one” being very confident for koala detection). The list of bounding boxes was transformed into single frames and heat maps that resulted in a binary image indicating where detections were found. These were computed to determine candidate koala signatures and then exported for manual review. The algorithm training went through an adaptive and repetitive process to fine-tune the results by using an initial model of true positives with previously verified data. The researchers applied it to the new data to identify false positives that were not a match to the model and annotated them as such. This allowed them to retrain the generic model, starting with the original ImageNet weights. The automatic and manual methods were evaluated independently and matched to verify accuracy and precision. This yielded a higher overall probability of detection of 87 percent for the proposed automatic method compared to the manual method, which resulted in a 63 percent probability of detection. The average time for automated detection of koalas in thermal footage from one RPAS survey was 136 minutes, with no operator needed. This average is shorter than the time of 170 minutes for manual detection. “On average, an expert koala spotter is going to get about 70 percent of koalas in a particular area,” Hamilton said. “We, on average, get around 86 percent. That’s a substantial increase in [the] accuracy that we need to help protect threatened species. Nobody else has really managed to get good results anywhere in the world in a habitat this complex and in these kinds of numbers.” With the success of the study, the researchers at QUT are looking to expand their area of investigation and target of detection. They hope to adapt this drone-based technology to examine endangered and invasive species.
