Bats may conjure up images of vampires, but forget the idea of them as shape-shifting, blood-sucking creatures of the night: Of the 1,400 bat species, only three actually drink blood.
Bats are not creatures to be feared, and they play an important role in nature as seed dispersers, plant pollinators, and pest control agents. It is estimated that bats’ consumption of pests saves the U.S. corn industry alone more than $1 billion in crop damage and pesticide costs each year. For all agricultural production, this figure adds up to more than $3 billion annually.
For bats to continue to play a vital role in our ecosystems, they need to be protected from a range of human-made threats, including habitat destruction, light pollution, wind turbines and forest fires.
That’s where the Bat Conservation Trust (BCT) comes in. Founded in 1990 as an umbrella organisation for more than 80 bat colonies across the UK, the charity works to support the recovery and resilience of bat populations, identify threats, protect habitats and dispel misconceptions.
Data collection and analysis is essential to carrying out this work. The BCT produces official statistics on the trends, abundance and locations of UK bats (such as horseshoe bats and oil bats).
Much of the Trust’s work is centred around acoustic monitoring, recording and analysis to understand which bat species are present in a particular area. Bat calls are usually at frequencies too high for humans to hear naturally, but can be listened to using bat detectors and recorded on SD cards.
The main detectors used by the Bat Conservation Trust are static, credit card-sized detectors that are around one-tenth the price of traditional bat detectors. The dramatic drop in price has enabled the charity to increase the number of detectors they use and the amount of data they collect from their citizen science surveys. Dr Lia Gilmour, Conservation Project Officer at the Bat Conservation Trust, explains:
A lot of our work is citizen science – we leverage citizen scientists to collect huge amounts of data and then use that data to ask questions about populations, monitor species status and make predictions for the future.
Sound and Vision
One night’s worth of audio recordings accounts for around 24-25 gigabytes of data, and with hundreds of detectors scattered across the country, it quickly becomes overwhelming. These vast amounts of data will enable the BCT to better understand the threats facing bats and how emerging global changes may affect different species. Gilmore adds:
As some bats move north, there may be a mismatch in availability of the insects and foods they need in their new locations, which is important to understand and address.
This is where the latest technology from AWS comes in. AWS supports auto-identification software that makes the process of analyzing all that data more efficient and faster.
Traditional data analysis is done manually, by first converting bat calls into images called spectrograms, and then analysing the images for features that can identify the bat species.
The AWS UK Imagine Grant has enabled BCT to develop a new version of its Sound Classification System (SCS) acoustic survey tool, which fully automates the audio identification of bat calls, removing the need for manual intervention.
SCS is hosted on AWS, which BCT chose because Martin Newman, IT research and development specialist at BCT, had worked with another client using the technology and felt it would be a great fit for the project.
Newman put together a proof of concept shortly before the COVID-19 pandemic last year. Since then, BCT has been working with partner Lambert Labs to make SCS truly cloud-native. Newman adds:
What I did wasn’t actually cloud native, but there were huge benefits in terms of efficiency, cost savings, and ease of use to being cloud native.
Lambert Labs developed a scalable, cost-effective audio processing pipeline for BCT, based on Newman’s existing proof of concept that ran on EC2 instances that could not be scaled and required reimplementation to scale to process over 100 terabytes of input data.
Lambert Labs also assisted BCT in its application for an AWS Imagine Grant to deploy a serverless technology stack to process large amounts of data without incurring high costs during idle periods.
One of the great things about serverless and AWS is that it scales very easily, so our SCS works for both relatively small citizen science projects and very large research projects – it can scale for both with equal ease.
It would not be economical if we had to buy a lot of computers to do the analysis: we collect data over a relatively short period of the year, which means we need a lot of processing power for a short period of time.
Timescale
BCT’s largest project last year collected 66 terabytes of data. By using serverless technology to send the data to AWS, BCT was able to process it in just 10 days. Newman explains:
To make that happen, we had to run 300,000 batch jobs. To listen to the recordings in a way that would allow for human identification, even at high frequencies, would have taken about 150 years of work because you can listen at lower frequencies. We went from 150 years to 10 days, and it all became doable. Technology helped us every step of the way to make this happen.
BCT has also been working with UCL and the University of Edinburgh to use machine learning to improve the classifier aspects of bat detection. The resulting BatDetect2 technology uses BCT’s training data to determine whether bat calls are present in an audio recording and identify specific bat species.
With new technology in place, the BCT will be able to tap into previously untapped information to discover more important details about bat populations and trends.
Previously, we relied on small amounts of data to make fairly large inferences about populations, but this could allow us to leverage information in ways we couldn’t before.
As an example, the BCT is currently looking into whether acoustics can be used to identify whether bats are breeding in particular locations.
If we could do that just by eavesdropping, without intervention or invasive techniques, that would be great, because it would tell us if there are maternity roosts or breeding hens, and therefore that the species is doing very well in that area. There are only so many questions that can be asked from this acoustic data, it’s just a matter of learning how to mine it and asking the right questions.
Next, BCT will look to incorporate a soundscape algorithm into the SCS, which will enable it to understand the soundscape of forest habitats and see if there is a ratio of artificial to natural sounds that indicates habitat health.
Our three-year vision, if we make it to the end, is to automate as much as we can to reduce staff costs. A lot of the things that AWS has started to do cost us a lot of money as we no longer have to do things manually. It’s not that expensive to run your data on AWS, it’s just a fraction of the cost, so the more automation we can do the better.
While the BCT is keen to automate as much of the bat monitoring and identification as possible, Newman points out that human involvement is still needed, including retraining the algorithms. He adds:
Machines are learning from us, but we need to be present too.
