AI and audio-identification of bats: Lia Gilmour interview part 1

Learn about BCT’s cutting-edge audio-identification work in this interview with our Research Manager, Dr Lia Gilmour. Lia talks about the incredible AI-powered acoustic survey tool – BCT's SCS - her team is developing. She explains how this system will remove barriers to surveying, improve bat conservation and increase impact on policy.

BCT’s SCS is an end-to-end audio-identification system, which basically means the whole survey process is done through it.

So, people can book equipment, upload files, get bat sounds identified and generate reports all in one place.

It is artificial intelligence (AI) that does the identification, though human users will verify them. And humans also train the AI.

Bat population monitoring is the grounded application of this cutting-edge science.

BCT produces the official statistics on the species, which are used by government and to inform conservation efforts. So, we need the best data. As well as building on monitoring data the BCT has going back into the 1990s, this technology will help us monitor a wider range of species.

Currently, BCT is at the forefront of this kind of research because we invest resources to shape its development. Overall, it’s vital that we're leading the conversation on how to monitor bats and develop these new techniques that are the future of monitoring.

We urgently need to upscale bat surveys, because bat species are ecological indicators. So as well as informing conservation and policy, our survey data lets us track ecosystem health. It’s increasingly important to do this in the face of global changes.

However, conventional bat surveys are time intensive. They need laborious field surveys and expert sound identification skills. Plus, although passive monitoring means we can generate survey data without using a lot of human time, we still need to process that data, which can be a huge task.

Using this grant, and the AI and cloud technologies that come with it, means we can now process that data with unprecedented efficiency and accuracy.

Unlike traditional surveys, SCS users don’t need to watch bats in real time. So it is accessible to those who can’t take part in physical surveys. Instead, the sensor can be put on a balcony or in a garden to collect audio data. It also means more people in towns and cities can take part, for example as part of the NightWatch and BBatS surveys. Identification is also ‘gamified’ to appeal to younger demographics.

Users get bespoke reports about bat species in their area, which helps connect people to nature, which we know is good for health and wellbeing. And some users can get specialised training to verify identifications, which will increase the number of people who can do this work. Once they have that training, they can then help fine-tune the algorithm. And this will expand EchoHub (an open-source library of bat sounds). Also, calls from EchoHub of good quality and known species can be used in turn to train the algorithm.

All of this will create gold standard datasets. And give people useful, transferable skills.

Traditional analysis of bat calls is manual. First, we convert a bat call into a picture, called a spectrogram. Someone then analyses that picture to see what characteristics it has. A long sweeping tail through the frequencies and then a bit that's more constant means it’s a pipistrelle.

You get your ID, but it’s time consuming, specialist work. And importantly, we don’t know what we miss because of our human preconceptions.

Auto-identification systems are software that looks at those bat calls and provides an identification rather than a human. When I started bat work, they weren't that accurate, but that's changing. The ID systems that we research and use at BCT use artificial intelligence based on neural networks from the brain, which we train with known species calls.

We train AI with datasets of bat calls, so it can recognise different calls in different parts of the picture. As we give them more data, the more AI will understand the huge variation that exists, even within species. And the more data you feed them the more they improve.

In the future, AI should be able to learn more than humans could about what's going on in the training data set.

However, there will always be a need for an expert human eye!

So much quicker! To give an idea, the data going through our system from this year alone would take a person a decade to analyse manually. But we have results in a matter of months.

What’s crucial to understand is that in some animals it's easy to say, ‘this sound means that animal’. But bats can be flexible. So, they may have a characteristic call type, but within that there's a huge range. And that range is dependent on the habitat, region and even what other bat species are present.

Plus, there's lots of untapped information in bat calls. We hope that new technology will give us access to ‘big bat data’ we can’t currently use.

Now, I must stress that it's always going to need human oversight. Humans need to make sure that the AI is trained properly and validate results. And importantly, to be make sure that AI is used in the right way.

Certain species may always need some targeted surveys. For example, Passive Acoustic Monitoring isn’t always possible for some species that have cryptic calls (calls which have a similar acoustic structure to other species’ calls).

But even then, AI can give a baseline before more intensive surveying. And I feel that there will be the nuances in cryptic calls that we can’t pick up with the with the human eye.