How ZSL is using Cloud AutoML to prevent poaching with automated image-tagging

The Google Cloud AutoML platform lets organisations such as the Zoological Society of London build bespoke machine learning models


The Zoological Society of London (ZSL) has turned to Google's new Cloud AutoML platform to track wildlife by automatically analysing millions of images captured by cameras in the wild.

These cameras help ZSL to conserve different species of wildlife by identifying their movements and potential poachers.

Image: iStock/SeppFriedhuber
Image: iStock/SeppFriedhuber

The motion-triggered camera traps use heat sensors to identify when wildlife or humans move past, and produce vast amounts of data that quickly needs to be tagged.

Reporting on that data typically takes up to nine months, by which time the animal movements and ZSL strategies may well have changed.

"You need that information much quicker in the fight to conserve wildlife or stop poaching," Sophie Maxwell, the conservation technology lead at ZSL tells Computerworld UK.

Google's Cloud AutoML uses artificial intelligence and machine learning algorithms to cut those nine months down to an instant. The platform helps organisations with limited machine learning expertise, such as ZSL, to build their own high-quality custom models using advanced machine learning techniques and tools.

Why AutoML?

Before turning to Google, ZSL conservationists had to manually tag these hundreds of thousands of images, a painstaking and time-consuming process. Cloud AutoML can do all this automatically.

It stood out from its competitors as it could use ZSL's bespoke models and train them on the charity's vast existing dataset.

Maxwell can't release results on the trials but says the potential was proven by early analysis of the images. Cloud AutoML has proven especially impressive on the tricky task of recognising species subsets.

"That's when this AutoML is much better than other image recognition solutions that are out there, because you're able to do these bespoke models based on your bespoke species set and your bespoke location," Maxwell says.

"It's not just more top level, so you can train it on your existing dataset and then you can start to define subspecies within that group, whether that's an impala or an oryx within this antelope family for example."

Read next: Best machine learning tools for data scientists and developers

ZSL's long-term goal is to create image recognition models that any conservationist can use to identify specific species and humans.

It hopes to embed the model in a chip that could be inserted into a camera on its Instant Detect monitoring system. ZSL could then scan images for potential poachers and send real-time alerts to rangers. Maxwell believes this only scratches the surface of the technology's potential.

"In the future it would be amazing if we could amass all this data to give a health check of the planet so that all this monitoring data could be collated, amassed and analysed to be able to give a more real-time view of what's actually happening out there, from pollution to species," she says.

"We're nowhere near that yet, because you need the data in the first place, but these steps are taking us there."

Technology in the wild

Cloud AutoML is a welcome addition to ZSL's tech stack. The charity's conservation projects depend on cutting-edge but affordable technology, such as facial recognition software to track known poachers.

The technology department supports a conservation programme team of around 300 people working in wildlife habitats from Africa to Antarctica.

These conservationists are often tasked with using old software on legacy systems due to budgetary constraints. This technology also has little commercial value, as it's designed for a small pool of conservationists by a charity that's funded through grants.

Despite these constraints ZSL has produced a series of ground-breaking innovations, including GPS animal tracking systems and data analytical tools that help spot signs of poachers.

The technology team either develops its own tools in-house or sources existing technologies that can be adapted to ZSL's needs.

Read next: How Callcredit uses machine learning to spot fraudsters and likely loan defaults

The in-house team of eight includes project managers, technical field specialists and software and mechanical engineers but the majority of its tech is developed with partners such a Google. The search giant has collaborated with ZSL on the development of the Cloud AutoML platform since early in 2017.

ZSL provided Google with around 1.5 million tagged images to help advance the platform. The charity is now training a custom model for its unique needs by feeding it data on conservation details such as region, environment and species.

"Once all of that is set up in the cloud it's very easy to connect to it via an API and then get predictions on unlabelled data that you get through," says Maxwell. 

"That's really our goal, to create these models bespoke for different regions and different species sets, whether that's Africa or Costa Rica or any of the other places that we work, so scientists and researchers can just upload their data and quickly get predictions back on what's in that data."

The insights will help them understand the location of the wildlife and its threats, in conjunction with acoustic monitoring and ground sensors. This information can help park rangers maximise their resources, by sending staff to the right spots in the huge areas they cover.

From the lab to the wild

Deploying the technology can be even harder than developing it.

It may need to send alerts about fishing vessels in the Indian Ocean to law enforcement on the coast, or of poachers spotted in the sub-Saharan African bush to a ranger station in a hut in the wilderness.

Connectivity or power is hard to find in these remote settings, and if it's there to power the devices there may be elephants around the corner who are all too happy to destroy them.

Read next: How Google Cloud IoT Core helps Smart Parking process vast volumes of data

They also have to contend with extreme environmental conditions.

"They have to cover the full spectrum of extreme conditions, from sub-zero temperatures in Antarctica to deep rain forests with high humidity, or environments where there's huge searing heat or low temperatures at night," says Maxwell.

"This is why there are so few commercial technologies for these areas. Getting open technologies at the right price for the conservationists that are accessible and democratised for all conservationists are our key aim."

The auto-tagging speed, ease of use and low cost of Cloud AutoML could make it a vital tool in ZSL's conservation projects.

"Recommended For You"

How Uber deploys machine learning-as-a-service with Michelango How technology helped foil the rhino poachers