NatureScot is the Scottish Government agency based in Inverness charged with protecting and enhancing the nation’s habitats and species. As a Statutory Consultee, NatureScot regularly provide expert advice on how Scotland’s environment and wildlife may be impacted by proposed developments.
The problem: camera traps produce thousands of images, but only a tiny proportion are relevant.
The solution: MSc student applies up to date knowledge of the latest developments in the field, develops a software model to detect, recognise and classify species and uses deep learning to train the model for accuracy.
It didn’t occur to us that Ioannis would be able to come up with an actual machine learning solution for species recognition in three months, but that’s exactly what he did. So it has far exceeded our expectations – we are delighted with what he’s done.
Information Management Programme Manager Stephen Gerrard, NatureScot
Data driven decision-making
A large part of the evidence used by NatureScot to inform these decision-making processes comes from their work on identifying habitats and the species present within them. Using an extensive network of motion-activated camera traps, NatureScot can gather primary intelligence on the range and concentration of various species, including protected animals such as the Scottish wildcat.
Information Management Programme Manager Stephen Gerrard oversees the management and sharing of such data:
Our camera traps produce thousands of images, but since they’re motion activated, this can include things like branches blowing in the breeze. Only a tiny proportion of the images we capture are actually relevant for our work, and we were having to manually trawl through them to identify the ones which were worth uploading to our Digital Asset Management system.
Stephen and his project board believed an Artificial Intelligence (AI) solution might exist to help them streamline their process of reviewing the camera trap images. With no AI expertise inhouse, Stephen, with the agreement of IT colleagues, applied for NatureScot to be a host organisation in The Data Lab’s MSc Placement Programme.
The Challenge: Image Detection and Recognition
Within hours, Stephen was matched with Ioannis Katsadas, a student on the MSc Data Science postgraduate programme at the University of Glasgow. After discussing the problem facing NatureScot, the pair agreed on a project scope which would see Ioannis research and assess various options and identify the best AI solution to reduce the amount of time members of staff would have to spend physically reviewing and classifying images before uploading them to the Digital Asset Management system.
Ioannis identified the need to apply deep learning, a subset of AI machine learning, to achieve the desired results. In order to correctly recognise and classify species within an image, the model would first have to be taught to detect the presence of the target species within the 2D images produced by the camera traps. Rather than manually annotate all possible outlines of each of the 50 species the programme might have to learn to recognise, Ioannis turned to open-source detection applications.
The detection capabilities in Microsoft’s AI for Earth project proved ideal for NatureScot’s purposes. The classification task, however, would need to be custom built, as Microsoft’s solution was based mainly on tropical species, and so less attuned to the highland and woodland habitats in Scotland!
Applying deep learning to the software model he created with fantastic results
Stephen had expected Ioannis to recommend a classification solution for later implementation. However, having done so with time to spare on his three month placement, Ioannis not only identified the best classification method but also created the model and left guidance on how NatureScot staff could use it in the future.
Creating custom script to pull images from external servers, Ioannis built a software model which could detect, recognise and classify species. Applying deep learning, he taught it to improve the accuracy of its predictions by correcting each mistaken classification until the model was ready for use.
Mutual benefits – while Ioannis developed valuable soft skills, NatureScot saved time and money
Within a few months, NatureScot had already filtered over 120,000 images through Ioannis’ model, freeing up over £20,000 worth of employee hours to devote to other projects and improving the quality of their image dataset. They also avoided the need to invest a higher amount with a specialist AI supplier, while benefiting from Ioannis’ deep up to date knowledge of the latest developments in the field thanks to his ongoing studies.
I saw how some things remain the same, for example the process of researching, looking at what has been done elsewhere, and what’s different, was similar to my academic experience. I found the big difference was in dealing with real-world challenges where things aren’t made simple for you. For example, we had problems accessing images on external servers and cleaning them to prepare a training dataset, so I had to create a special script to prepare the data.
Drawing on his problem-solving experiences, as well as the soft skills he picked up in collaborating with team members and presenting his work to the wider organisation, has already helped Ioannis differentiate himself in a competitive job market. He believes the dissertation he produced based on the work he did during his time at NatureScot was more fulfilling than a purely theoretical piece, and has since graduated with a distinction from his MSc in Data Science.
The remote project went very smoothly, with support from The Data Lab
With the 2020 programme taking place while the world grappled with COVID-19, Stephen and Ioannis found themselves collaborating remotely, rather than from NatureScot’s office as originally planned. Having returned home to Thessaloniki in Greece, Ioannis worked as a virtual member of the team. We were arranging the placement just as lockdown was starting,” remembers Stephen, “and we worried about whether it would be an extra layer of hassle, and whether we should even go ahead.”
With NatureScot already geared up for remote working, however, they realised the project was perfectly suited to this and went ahead, making sure to involve Ioannis in wider team videocalls and virtual get-togethers. “It was really smooth and easy. It was all far less work than you would think.”
The Data Lab was also on hand to support both student and sponsor, particularly as it was NatureScot’s first experience of being a host organisation in the MSc Placement Programme. “We did a lot of work with The Data Lab at the start,” says Stephen, “Marian was very helpful, and Karen too. They talked us through the process and also had regular catchups with Ioannis and myself, both individually and together.”
If you’ve identified a data-based business problem you’re trying to solve, I’d definitely recommend getting in touch with The Data Lab to see if they might be able to help.