Home PublicationsData Innovators 5 Q’s for Peter Kohler, Founder of The Plastic Tide

5 Q’s for Peter Kohler, Founder of The Plastic Tide

by Nick Wallace

The Center for Data Innovation spoke to Peter Kohler, founder and chief executive officer of The Plastic Tide, a UK-based charity that is developing an algorithm that can recognize plastic waste on beaches in images from aerial drones. Kohler talked about how aerial photography can help beach clean-ups and improve understanding of the global problem of marine pollution.

This interview has been edited for clarity

Nick Wallace: What kind of algorithm is The Plastic Tide training, and why?

Peter Kohler: Essentially, we’re looking to create a digital eyes and brain for monitoring marine litter and plastics. The ocean is under siege with human-produced waste, yet we don’t know to what extent—we don’t have a clear picture of that. We also don’t know where the concentrations are, or what impact it’s having on marine life, and also human life as well. There’s a big interaction between humans and the sea, as well as marine life. Current methods of measuring and monitoring are human-based, and they are variable by time and space. There’s a lot of uncertainty built into the data, so scientists think the estimates may be underestimates, and there may be ten times more plastic on our beaches than we think.

What we are doing is building a scalable, cost-effective system that could be applied globally to monitor plastic and marine litter, using the machine learning element to automate and scale that in an affordable way. It could be anything from citizen scientists with drones going down to the beach and using the machine learning algorithm, all the way up to deploying autonomous platforms at governmental level, monitoring whole coastlines with autonomous drone systems. Machine learning, especially in the last couple of years, has offered that ability—it wouldn’t have been possible a few years ago.

You can help train our algorithm right now as we talk, if you want. We upload the images so you can look at them and identify pieces by drawing boxes around them, and classify them from a very basic list. That provides tags for the machine learning algorithm to learn from. We’ve had over 10,000 people donate tagged data, and it’s built a database of over 4 million classifications. That’s the order of magnitude you need to train a machine learning algorithm on something as diverse and as complex as litter.

It’s the same sort of technology you would use for driverless cars, where a human is a basic shape—there’s a head, usually four limbs, a torso—the algorithm can learn to distinguish that quite well. But the problem with plastics and litter is that it’s all sorts of textures, materials, colors, shapes, sizes, so you really need a much bigger sample dataset to train in.

That’s what the public are doing now—they’re training the algorithm. There’s basically almost no other way to do it really, because either way you look at it, you need the training dataset, and what that develops is a baseline. It’s the only dataset in existence in the world today for box-wise localization—where you’re drawing a little box around the plastics. 

Wallace: What happens once you get to a stage where the algorithm can reliably identify plastics on beaches? Where do you plan to take the project next?

Kohler: What we want to do is reveal and relieve the siege of waste that our oceans are under at the moment. We reveal it initially by generating all the information about plastics and marine litter. In the short term that will mean we will be able to coordinate cleanups much more effectively. We’ll be able to say, “if you’ve only a certain amount of time on the beach to clean, this place is the best place to go.”

In the medium to long term, by informing the policy cycle, by a process of monitoring, evaluation, research, and learning, we can build that scientific dataset for enforcement, for policy, and for campaigns. We’ll be able to look back and say, “what has that actually done for the environment?” Has it reduced plastic bottles washing up? That’s when we can get to the stage of “turning off the taps” on land, as it were, by seeing where the problem is concentrated and where it’s building up.

At the moment, we don’t know. There’s no mechanism to report how much waste people are dumping into the sea—not that anyone would report that anyway. The only way is to monitor the inputs into the ocean by using an algorithm. In the short term it’s damage mitigation, in the long term it’s prevention, through policy enforcement and campaign-based input.

Wallace: Besides making it a bit easier to clean up beaches and informing policymakers, what impact do you think this kind of data collection could have on the problem of sea pollution? Is this just about looking at plastic on beaches, or is there more to it than that?

Kohler: A big part of what we do is also behavior change. There’s collecting all of this information—that’s the revealing part I mentioned. But it’s also about empowering them to change behaviors. That is the crux of the problem: behavior change on an individual level and all the way up to a corporate level and a policy level.

It’s similar to climate change: we’ve got used to a certain way of doing things. I always use the oil tanker analogy: it takes a lot of energy just to shift an oil tanker by a few degrees, for it to stop and change direction. That’s essentially what we’ve got to do. So what I see us doing, as well as collecting all the data and providing the policy, is also using that to educate and to empower people to get involved. We’ve had a lot of anecdotal evidence to say that works. People are tagging all the images, and they start looking for it, they start to see the plastic as they walk down the street. Then they become a lot more aware that there is litter in the environment, whereas before they tuned it out. Really my dream would be to get people to not consider the environment as a separate entity, but as one and the same as them. Again, that’s personal—you’re not separate from the environment—but it also goes up to the corporate level.

Wallace: What made you look at AI as a way to tackle this kind of pollution, and how did The Plastic Tide get started?

Kohler: I’ve always been passionate about the oceans. I was actually born in Kathmandu in Nepal, which is 1,400 meters above sea level—it’s one of the highest capitals in the world, so you can’t really get much further from the sea than that. But I would go to Thailand on holiday and see this blue ribbon, and the expanse and the space, which would fascinate me for decades.

I studied geography, and then I went traveling in 2008, and I lucked-out with a skipper who was looking for crew. I jumped on board and sailed the South Pacific for four months. And it was paradise, but it was one that was under siege. I could see litter and pollution—there are island nations where waste management is a concept that’s only recently been introduced. But there would also be litter in places where probably the nearest people would be a plane or a space station flying overhead. Yet there was still massive bunches of litter on the beaches, so I wondered, “where is it all coming from? If we’re so far away from civilization, how can all this litter be piling up on this beach when there’s basically no one here?”

So then I came back to the UK, I continued my day job, I attended all the ocean events I could possibly attend, and researched the issue. I hit upon the fact that 99 percent of ocean plastic is missing— this is the active area of research of Dr. Erik van Sebille, one of our science advisors, and he’s currently building up an estimation of that based on papers, but there’s no real statistical hard evidence to say where that plastic goes.

Then over a pint with a couple of mates, we were trying to come up with an idea to win some funding, or an adventure, and I had the idea of driving the West African coast and collecting litter as I went, and surveying it for science. But then I thought, “well, jeez, it’s going to take me a long time to do that manually, surely I could use drones to image that!” I got to speaking with a couple of friends, one of whom is Dirk Gorissen, one of our trustees, and he said “yeah, there’s this thing called machine learning, and you can apply that to this problem.”

So it just snowballed from there. I met a couple of scientists, they really loved the project. We got together, we had a master’s student volunteer his time, and then we very quickly, in less than four months, executed 33 beach surveys—and beach cleans, which I do not advise you to do if you want to maintain your sanity. That led to the database, which was then tagged—and that’s where we are now. We lab-tested a prototype back in July, and now we are looking at doing a trial on a beach using the algorithm. We had another master’s student work on it, which has improved the accuracy considerably, and we’re about to go public with how much that is—but we need to test it in the right environment to say scientifically that we have that level of accuracy.

We also sampled the UK coastline in April and May last year. That was to gather the base dataset of images, which was something like 14,000 images. Most of them have now all been tagged. Following on from The Plastic Tide I got contacted by a lot of drone operators from all around the world—lots of them. So I set up what we now all Blue Earth Litter Surveys, which used to be called the Marine Litter Drone Network, that’s a separate entity from The Plastic Tide intended to develop a standard approach to drone surveying—aerial, submarine, and sea-surface. As part of that, some of the members have got funding by partnering with us to survey their beaches, and they give us the images. At the moment, that’s how we’re working. The images you see on our tagging platform now are from New Jersey in the United States—those are from Morris, he’s one of our members at Blue Earth Litter Surveys. So we’re still getting a slow trickle of images in.

Wallace: Are you using any other data sources, like satellite images?

Kohler: On Tuesday we’re going down to the south coast of England, and we’re going to fly with some infrared spectrometry and compare the readings with our visible light imagery. Deploying infrared imaging of plastics in the environment is quite an experimental area, as far as I’m aware. Combining that with machine learning and visible light elements, and we’re quite interested in how we can merge the two.

At the moment the satellite resolution is a really limiting factor, but in ten years, freely or commercially available imagery could get down to the resolutions that we need, which is centimeter level. At the moment, the only way to get that is with drones or aircraft. But for sure, we see in the future being able to monitor whole coast lines, potentially with satellite imagery. That would just be a case of some retraining of the algorithm on the satellite imagery—if the resolution is sufficient, it should be similar to the visible light.

One of the challenges that we face is the manual surveys have something like 100 classifications for plastic litter. Asking people to choose from that list when tagging the images is just not possible. So we have very basic categories at the moment. That’s where we hope the infrared element would be applied. If we can identify the polymer types, we can identify the plastics. But even then, if you introduce a red dye to your plastic, you can potentially affect the spectral signature of that plastic, and the different combinations could be in the millions—so that’s an interesting research angle.

The AI could help in doing that. But at the moment, what I understand is that the algorithms are not yet capable of picking up that information without having been taught it first. Maybe in the future, an algorithm will be able to teach itself from other sources. You could set it a task to trawl the Internet for mentions of litter and pollution—but that requires a semantic understanding of language, because when people say “that’s a load of rubbish,” they’re not actually referring to rubbish. So the algorithm would have to understand that. But at the moment, for the foreseeable future, this tagged database is the fastest way of training machine learning for our purposes.

Image credit: David Altabev

You may also like

Show Buttons
Hide Buttons