Turning camera traps into real-time sentinels: Interview with Conservation X Lab’s Dante Wasmuht

Camera traps are ubiquitous in conservation. They’re deployed to monitor biodiversity, study animal behavior, observe habitats over long periods of time, and enforce effective conservation action on the ground.

However, they have an inherent shortcoming that’s especially prominent in remote locations like dense forests and cut-off islands.

“They’re largely passive. People put them out and then half a year later, or a year later, they go and collect this,” Dante Wasmuht, head of artificial intelligence at the nonprofit Conservation X Labs, tells Mongabay in a video interview. “From the moment something happens in that area until a human knows about it, a lot of time can pass.”

To address this issue and encourage early intervention, Conservation X Labs has developed an AI-powered tool that can be attached to camera traps to provide real-time updates on what’s happening in the forests.

“It’s as if that camera is connected to a small computer which is the Sentinel,” Wasmuht says. “Whenever the camera trap takes a picture, the Sentinel knows and transfers the image onto its inner computer chip and can run all those AI models completely locally right there in the field.”

The Sentinel has been deployed in various countries to track invasive species, monitor wildlife poaching, and keep an eye on threatened animals.

In New Zealand, more than 80 Sentinel devices have been used to monitor invasive species on remote islands. On Ulva Island, for example, the device played a critical role in detecting invasive rats that were wreaking havoc on the ecosystem. The data helped conservationists enforce intervention measures before further damage was wrought. The device has also been used to track invasive Burmese pythons in the Florida Everglades.

However, challenges continue to exist, especially with regard to weak connectivity in remote areas and the lack of reference data to identify rare species.

“There are species that are not present in any sort of existing AI models,” Wasmuht says. “How are you going to make an AI model that detects the species if only 10 images of that species exist?”

Dante Wasmuht speaks to Mongabay’s Abhishyant Kidangoor about how Sentinel works, why it’s needed now more than ever before, and the challenges in using it on the ground. The following interview has been lightly edited for length and clarity.

Mongabay: How did you get into the field of conservation technology? What is it about using AI for conservation that excited you?

Dante Wasmuht: I’ve always been very passionate about conservation and nature. I spent time in Madagascar and have been interested in wildlife and nature in general.

My career was in neuroscience. I did a Ph.D. in neuroscience, more focused on computational neuroscience. Now there’s a lot of hype around AI and all the things that it can do. But if you look a little bit at the history of AI and specifically neural networks, there’s a very close relationship between neuroscience and AI. Those algorithms are called neural networks and they, in fact, do take some inspiration from the brain. So I worked a lot with that during my Ph.D. and got really in touch with AI. When I finished my Ph.D., I came across Alex [Deghan], who’s the founder of Conservation X Labs, who wanted to apply technology. To me, that was somehow a very good marriage of my passions and background. That’s how I ended up working in AI for conservation.

Ten to 15 years ago, there was no AI for conservation. Nobody was applying artificial intelligence to anything relating to conservation. I think that whole field really exploded simply because there’s so much data that’s being generated and there are so many sensors that are being put out in the world to measure things and to monitor things. AI is probably the only way to deal with that much data and make sense of it.

Mongabay: How would you describe Sentinel to someone who hasn’t heard about it?

Dante Wasmuht: Basically, Sentinel is something that connects to your existing camera traps. It’s not like a completely new device or anything. It’s more like an adapter that you can use to retrofit an existing camera trap. It’s as if that camera is connected to a small computer, which is Sentinel. Whenever the camera trap takes a picture, Sentinel knows and transfers the image onto its inner computer chip and can run all those AI models over the data completely locally, right there in the field. Then, what has been an image of something becomes some sort of insight or a learning. Like, there was a white boar on that image, or there were two types of species on that image, or that image was empty.

The advantage of this is that, first of all, you can filter out all the empty images and you can classify all the different animal species. Then you can compress this information and send this over very low-bandwidth communication areas to a user dashboard or to people’s phones or by email. The big advantage of this is that people know right away what’s happening.

With the AI running locally, you don’t even have to send back all the raw data. You send back the super-compressed information, and that really allows you to work in areas where you normally don’t have any connectivity.

Mongabay: What gaps were you trying to fill when you developed this tool?

Dante Wasmuht: One of the main tools in all of ecology and wildlife conservation is camera traps. People put them out in national parks or wildlife corridors or very biodiversity-rich areas, and it’s basically a motion-activated camera that runs on batteries, and that’s supposed to be in the field for some time. Whenever an animal walks past, it activates that camera and it takes some pictures or videos.

The problem with camera traps is that they’re largely passive. People put them out somewhere and kind of forget about it. Then half a year later, or a year later, they go and collect this. So from the moment where something happens in that area until a human knows about it, a lot of time can pass.

Camera traps also generate a lot of data. A given camera trap can generate thousands and thousands of images if you leave it out for a couple of months or so. And normally these are deployed in large grids of 100 camera traps. That can easily generate a million images or videos in a year or more. Then you have to go and look through those. Normally, up to 90% of that data that’s being collected by those camera traps is empty. There’s nothing on it because the wind moves the vegetation and it gets triggered.

So we have two problems there. You have massive amounts of data that you have to process that it’s almost impossible for humans to go through. This is compounded by the fact that action timelines are extremely long because first you have to go get the data, and then you have to go analyze it. This is clearly a perfect use case for something like AI.

Mongabay: Why is this technology more important than ever at this point in time?

Dante Wasmuht: You can’t act on what you don’t know. I think what is specifically important about Sentinel is this shortening the data to action timelines. This is super important for things like antipoaching. You don’t care whether there was a poacher a year ago. You want to know whether there is a poacher now so you can do something. It’s the same for lots of endangered or threatened species. Many times, you want to know what happens relatively fast so you can do something about it.

Another major use case of Sentinel is detecting and managing invasive species. I’m talking about invasive mammals and reptiles, something that you would normally see on the camera trap as well. They do cause enormous damages. It’s really sort of something that turns a passive, long-term monitoring effort into some sort of real-time monitoring system that you can actively use as management.

Mongabay: Could you give me a brief history of the project?

Dante Wasmuht: We started out five years ago. We develop technology at Conservation X Labs and we do work with a lot of field programs. We have boots on the ground. And there are a lot of people who came to us and said that they didn’t know what to do with all this data.

As we know, conservation is historically underfunded. It’s small NGOs that are trying to do big things. Usually, you have a small team of people that try to manage 40,000 hectares [about 100,000 acres]. It’s completely infeasible to set up something like a camera trap because you need just weeks to set it up, to collect the data and to analyze the data. That’s almost impossible to do for a small team.

That’s why we really wanted to develop a solution that could act as a force multiplier so that you could actually go out and attempt a monitoring effort or invasive species removal effort. This really allows you to manage a relatively large grid with a few people.

Mongabay: Could you walk me through the tech? How does it work?

Dante Wasmuht: Sentinel is a little computer with an AI chip and communication modules. It can communicate to cell towers, to satellites, and over radio frequencies.

Whenever a camera trap gets activated, Sentinel runs the AI over it. The AI models have a part that detects animals in an image. They have a part that classifies those animals. And then they have something that compresses all that information so you send it back over those communications channels to your management platform.

There’s something that’s called a MegaDetector, which detects animals. MegaDetector has been trained on so much data that it generally works pretty much out of the box. So you can put it on Sentinel and it will detect animals everywhere.

Then you have classifier models. They have a couple of hundred species, or maybe 1,000 species. But you have to be careful because you want to work with the local teams so that you make sure that those AI models work well on local data.

We have big projects in New Zealand. The New Zealand Department of Conservation has collected millions and millions and millions of camera trap images over the years. They’ve classified them and labeled them. They know all the species they have, and we can use this to train a model. So we train the New Zealand species classifier with a focus on invasive species.

Sentinel is very flexible and so you can put whatever model on Sentinel. Sometimes we have models that are very general. Some have, like, 100 different species and they work in different places. But then we also have very specific models that work super well in New Zealand with 30-40 species that are very important in New Zealand from a conservation perspective.

Mongabay: How would you describe the impact of Sentinel?

Dante Wasmuht: We work very closely with the Department of Conservation in New Zealand, specifically with the team that’s responsible for island invasions. We have Sentinels on six different islands.

There’s this specific island, Ova Island. It’s very protected and it’s supposed to be rat-free. We deployed around 30 sentinels on that island. Over six months, we collected around 100,000 or so images and it turns out there were 45 rat events. There were actually rats on the island — rats that swam over from the mainland due to warmer water temperatures.

Then they could react to this. They deployed some sort of traps. But thanks to Sentinel, they knew in real time where the rats were moving and they could very specifically place those traps to catch the rats. If they weren’t using Sentinel, they would only know about the rat incursion whenever a ranger checks on all the cameras and analyzes all the images, which would take months. It really saved them management costs. It also saved them from costs that would go with large-scale eradication projects. In many cases, they would drop aerial toxins on that island. It’s like a super-aggressive strategy that kind of exemplifies how bad the problem is.

We have similar use cases replicated in New Zealand where we’re also looking at invasive possums and other species. We’re also looking at threatened kiwi birds and how they react to those eradication efforts, because they’re always directly coupled with changing populations of endemic species.

We’re still a little bit removed from the sort of impact where we could say, “Oh, we actually brought back X, Y and Z species.” But we are at the stage where we can say, “We detected X, Y and Z invasive species and decreased the cost of eradicating them.”

We have around 500 units deployed in 14 countries. I think our biggest deployment was 100 units on a 20,000-hectare [50,000-acre] island.

Mongabay: What have been the challenges in doing this?

Dante Wasmuht: The AI piece isn’t always easy. Because sometimes there are species that are not present in any sort of existing AI models, and there also doesn’t really exist data for those species. So that’s really challenging, because how are you going to make an AI model that detects the species if only 10 images of that species exist?

It’s a sort of computational data problem that we face many times. Then there’s problems with communications. We’ve been working in places and islands that are super remote. The only communication is through satellites. But then again, certain satellite providers don’t work in certain countries because of political reasons.

The other part that’s tricky is that you have to make this cheap. It’s something that we want to make affordable for conservation. It’s one thing to put 10 devices out, but for some of those projects, you need to put 200 devices. And that’s tricky, because we’re an NGO. We are not a big company.

You need a scaled solution to do this, and AI is the key to scale on the technological side. But then there also need to be incentives to scale from an economic point of view, which is tough.

Mongabay: What are the next steps with Sentinel? How are you thinking ahead?

Dante Wasmuht: The major next step is to scale this. We want to make this as cheap as a normal camera trap.

We would want to also venture into other modalities because this is an adapter that clocks into anything. It can also plug into acoustic recorders. Or it could be something that works underwater. That’s another sort of avenue. So yes, scaling it, making it cheaper and more user-friendly.

Invasive species monitoring is our main use case. But we would definitely also be stronger on the biodiversity monitoring side. The question always with biodiversity is: do you need real time? For invasive species, you need real time. For antipoaching, you need real time. For human-wildlife conflict, you need real time. But general biodiversity monitoring is a little bit tricky because this is slightly more expensive because of the real-time aspect and data transmission aspects. I think a big goal is to make it cheap enough that it’s not much more expensive to deploy than something completely manual or completely passive.

Imagine you have those devices that send all the information as processed by the models to this platform. All the species trends, all the populations, how it moves and changes in real time. Then you would only need to basically maintain this every now and then, and swap out batteries and things like that. But your job is really focused on the management aspect.

Banner image: A mountain lion (Puma cancolor) caught on camera trap caching a deer for eating later. Image by Jon Nelson via Flickr (CC BY-NC-SA 2.0).

Abhishyant Kidangoor is a staff writer at Mongabay. Find him on 𝕏 @AbhishyantPK.

Credits

Abhishyantkidangoor Editor

This story first appeared on Mongabay

This article is licensed under a Creative Commons Attribution-NoDerivatives 4.0 International License.

You may republish this article, so long as you credit the authors and Mongabay, and do not change the text. Please include a link back to the original article.