Emily A. Jordan discusses the use of AI in population monitoring and her team’s experience using it to assess the Kapitia skink.
In population monitoring, using unique markings to identify individuals is a practical solution when species are challenging to tag. We can camera trap elusive snow leopards, drone-photograph whales, and happily snap our tiniest amphibians. Yet these photographic records bring a fresh challenge. Each observation needs to be matched to a known individual – or assigned as a new one – based on our assessment of their features, and as the number of records grow this becomes increasingly difficult and time-consuming leaving us fatigued and prone to error.
Cue artificial intelligence, or AI. First coined in the 1950s, today AI can feel like an inescapable buzzword, but behind the buzz lies a sensible tool for practitioners. The term most basically refers to a computer performing a task associated with human intelligence, such as pattern recognition and language comprehension. Its scope can vary from huge generative models such as ChatGPT, to focussed software for repetitive tasks.
How does it work?
Most of us have probably used an AI phone app to help identify an unknown species; individual recognition software is similar – it just focuses on distinguishing features at the individual level. Several individual ID packages are available and, although they vary slightly in their make-up and user interface, they ultimately follow the same pipeline:
- Pre-processing stage – the user selects an area of interest for the AI to focus on (this might mean cropping, or highlighting patterns).
- AI compares the images – the algorithm works to score pairs of images based on their similarities.
- Visual comparison – the user is presented with a refined selection of images to accept or reject as matching pairs.
Does it make a difference?
So far studies suggest that using an AI package for identifying individuals improves accuracy relative to attempting to match individuals by human eye alone. As individual encounter histories are used to make population estimates, mistakes made at the ID stage are likely to ripple through to our estimates, something simulation studies already indicate.
We wanted to illustrate what this could mean in a real-world conservation scenario by looking at the effect of using an ID software (here we used HotSpotter) on the bottom line of demographic estimates for the Kapitia skink – a critically endangered lizard in need of urgent conservation. Comparing the outputs when analysing two distinct data sets; one constructed with the help of AI software, one with identification made by human effort alone, we saw notable differences in demographic estimates. In one instance, our results indicated that using AI prevented us from overestimating population size by more than 10% – highly significant when planning interventions for a species on the edge.
What’s the catch?
Whilst we feel that this tool is a great way for projects to minimise error in their monitoring, there are issues to be mindful of. Of the software available, most are free and receive different levels of maintenance (if any). It’s worth playing with a few options to see what works best for both your species of interest and computing system. Paid for versions can be smoother and can even automate the pre-processing and matching stages to a greater degree – but of course, there are costs. Finally, there’s still room for errors to creep in with this method – there’s only so much that can be done with a blurry photo – but ultimately if you have a monitoring project reliant on sorting through heaps images, an AI solution may just be worth a second look.
Interested in reading the full article in Ecological Solutions and Evidence? Please click here.
