Dugongs, Drones and Artificial Intelligence
You probably don’t think about seagrass very often...
Indeed, diving amongst seagrass is probably not on your bucket list, and you probably don’t have beautiful photos of seagrass adorning your walls at home. Even I, as a researcher of dugongs, an animal that totally relies on seagrass for their existence, have been guilty of underestimating the importance of seagrass. But seagrass ecosystems support 50 per cent of the world’s fisheries, making it a critical ecosystem for the livelihoods of three billion people who depend on these fisheries for food and income. What is perhaps even more surprising is the importance of seagrass in protecting us against global warming: seagrass can absorb and store carbon 35 times more efficiently than rainforests. Whilst we have been desperately trying to hang on to the last of our rainforests, seagrass is disappearing at a rate of two football fields every hour. We have already lost an estimated 35 percent of this precious resource.
Protecting seagrass, however, is difficult if we don’t know where it is. Seagrass expands across many millions of hectares and a large proportion of seagrass is too deep in the water column to be seen from a boat or from the air. So keeping track of how healthy our seagrass beds are is also difficult. One answer lies in monitoring the animals that rely on seagrass. And this is where the animal that has been the focus of my career comes in - the dugong.
Dugongs can be found throughout the Indo-Pacific, in a total of 46 countries. Their diet is made up almost exclusively of seagrass, hence they are also called the sea ‘cow’. So wherever you find dugongs, you are also likely to find seagrass. And, importantly, if your dugong population disappears, you’d best start investigating your seagrass, because that’s likely gone as well, and your fisheries will quickly follow suit.
But dugongs can be tricky creatures to monitor, as I quickly found out back in 2001 when I first sat on a boat and tried to record what they were doing. Their habit of feeding on seagrass means they spend most of their time on the sea floor. They are mammals, so regularly come up for air, but even then they tend to be quite cryptic, often just sticking their two nostrils on the end of their very strangely shaped head out of the water and then sinking back down to the bottom.
The answer is to observe these animals from the air. The aerial perspective allows you to look down through the water and see dugongs when they are still submerged, and in shallow clear water, even when they are feeding on the seagrass. Since the 1980’s, my esteemed colleague (and PhD supervisor) Prof. Helene Marsh has been conducting aerial surveys of dugongs in Australia, and she taught me everything I know.
Aerial surveys are a really efficient way of finding out where dugongs are, how many there are, whether their numbers are increasing or decreasing, whether they are breeding (how many calves there are) and of course, where their important habitats are, including seagrass. Dugongs occur along vast lengths of the northern Australian coastline, and these aerial surveys can efficiently cover hundreds of kilometres. So surveys have been a very effective way of keeping track of our dugongs and understanding what areas of habitat are critical for us to protect.
But hiring planes and waiting for suitably low winds for surveys is very expensive, so we have to be conservative about how often we conduct them in Australia. And there are only three other countries out of the 46 in the dugong’s range, where surveys have been conducted regularly. There are two other key reasons for this, besides the expense – aerial surveys are dangerous, and they require teams of experts. Normally there would be a team of five of us – a survey leader plus four observers – seated in a six-seater plane with a pilot, flying a set flight path consisting of many parallel lines stretching perpendicular from the coast. The observers call out the dugongs as they see them, but get a only a few seconds to see and distinguish the dugongs from other marine animals, so they need to be well trained and experienced.
We fly these surveys at low altitude (only 152 metres above the sea surface) and as slow as the aircraft can go, which doesn’t leave much room for error. Around the world, there have been five fatal plane crashes that have killed 11 marine mammal researchers. Many of us, including me, have had ‘incidents’ that have made the dangers of this research technique all too real. Losing an engine in a small plane and having to make an emergency landing can be a life defining experience. And I have a colleague (who also happens to be my husband) who has been in this situation three times!
So having conducted hundreds of hours of surveys, there was plenty of motivation for me to run with a colleague's suggestion, 10 years ago, to use drones. I thought it’d be easy and that I’d be able to come up with this new method within a year. Well… Back then, drones were not so common, so it was hard to find one capable just of flying the distances we needed to fly. And it was even harder to get aviation approval to fly beyond visual line of sight (the distance at which you can still see the drone). So developing this technique has taken many more years than I anticipated.
I was extremely fortunate to team up with Insitu Pacific Ltd, who provided in-kind support for my research by flying their amazing, high-end, military drone, the ScanEagle. We needed to understand how best to capture large areas of water in the aerial images from the drone while still being able to see the dugongs in the images. And then we had to compare how the number of dugongs sighted in the drone images compared to the number observers could normally see from a plane.
The result: the ScanEagle images provided data that was just as good as, and in some scenarios, better than, the data from the observers. So great, we can use drones now right? Well, there were still major impediments to switching to this new technology – the image processing and the cost of these high-end drones. One drone survey results in tens of thousands of images, which take hundreds of hours to manually review to find the dugongs.
Cue Dr Frederic Maire, from the Queensland University of Technology, a specialist in artificial intelligence (AI) and image processing. Frederic designed the ‘Dugong Detector’ – a computer algorithm that uses a type of artificial intelligence (AI) called machine learning to find the dugongs in the drone images automatically. The Dugong Detector shows the potential dugongs as small thumbnail images, which we then verify by marking the correct ones. We can then use the true and false detections to ‘retrain’ the Dugong Detector so that it gets better and better at knowing what a dugong looks like.
We have also teamed up with Martin Weiser from the Vienna University of Technology, who has written a program that can map all the drone images, using information recorded by the drone about its position and rotations. The software also takes the sightings identified by the Dugong Detector and maps every individual dugong. So we end up with very accurate information about the area we surveyed and exactly where the dugongs were.
You may now be thinking "but what about the dugongs that were not visible from the air?" Of course, there is a large proportion of dugongs that are too deep in the water column to be visible when we fly over. We have done a lot of research to understand exactly what this proportion is, and it is largely dependent of the water depth and clarity, and the wind chop on the water. So we need to measure these environmental conditions as well. Traditionally, these conditions have been assessed by the observers. But our next challenge is to use AI to automatically classify the drone images for the environmental conditions as well. We also want to record any visible seagrass and adapt the Dugong Detector to find other seagrass-reliant species, such as turtles.
Another remaining challenge is to understand what information we can get from small, cheap drones that are in the realistic budget of researchers working in the countries throughout the dugong’s range, which are mostly developing or under-developed countries. This is the mission of Dr Christophe Cleguer, at Murdoch University. Chris is working with two different-sized drones, the Phantom and the Wingtra, which have different capabilities and can cover different sized areas. We need to know what questions these drones can answer.
Finally we need to make these new survey techniques available to all 46 countries. We’d like non-experts to be able to take a drone, and do a dugong survey. To do that, they need access to advice about how to design their survey and how to select the appropriate drone. Then once they’ve completed their survey, they need the Dugong Detector to process their images and record the dugongs, turtles and environmental data. And they need our mapping software so they can map their dugongs and visible seagrass. Ultimately, we want to create a global database so we can have an understanding of the status of dugongs in all 46 countries worldwide.
So our mission is now to make all of these tools available on a Dugong Detector Online Hub. Armed with these tools, all ‘dugong countries’ will be able to gain an understanding of not only their dugong populations, but subsequently, their seagrass. This information will allow local communities and governments to better protect the vital seagrass ecosystems upon which their livelihoods depend. Luckily, our mission moved a step closer to achieving its aims when it won a 2018, AU$250,000 Google Impact Challenge Award. Watch this space!