Image Recognition and AI for Species Identification

This presentation will explore how image recognition and artificial intelligence (AI) can revolutionize the identification of insect species. By leveraging advanced machine learning algorithms and high-resolution imaging technology, UK CEH aims to enhance the accuracy and efficiency of monitoring insect populations. Alba will discuss the development and application of the Automated Moth Identification (AMI) system, which exemplifies these technologies in action. Additionally, she will highlight the role of citizen science in this initiative, enabling broader participation and data collection. This approach not only provides critical data for addressing the biodiversity crisis but also bridges the gap between technological innovation and ecological research.

Alba Gomez Segura is a Research Software Engineer at the UK Centre for Ecology and Hydrology, where she develops and applies new technologies to monitor biodiversity. With a background in Environmental Science and Bioinformatics, she now focuses primarily on electronics and coding. Her interests include hardware (e.g., cameras and Raspberry Pi) and software (machine learning) to enhance insect monitoring in the field, aiming to provide more and better data to tackle the biodiversity crisis. She is passionate about bridging the gap between technology and biological research, driven by a deep passion for understanding the natural world.

Q&A with Alba Gomez Segura

Where were the training images for the moth image recognition algorithm sourced from?
Our main source of images was the Global Biodiversity Information Facility (GBIF). As the model progressed, we were able to tweak the model using images gathered through the AMI, as GBIF images are often good quality and it is useful to add in images that are more representative of what the AMI will capture. With two seasons of AMI data from some sites, we now have lots of images that can be used for training. However, there are some limitations with this data such as the fact that some species may not come to the AMI very often or event at all. Therefore, we will continue to use both GBIF and AMI images for training the model.
Do you stop training the AI once you have reached the point where the accuracy is no longer improving?
We have a maximum number of images to help us focus our efforts on species that we have fewer images for. However, we can use more than 200 images if needs be and this may be particularly useful for species that are more difficult to identify. The number at which the accuracy starts to plateau is variable by species. We tend to tell people that at least 100 images are needed to start with, but it can be less for particularly distinctive species and more for species that are more difficult to distinguish from other species.
Are the images that you are gathering available through an open-source platform?
The idea is that these images will be stored on an open-source platform. The images that we have taken from GBIF are not owned by UK CEH and are already open source. For the UK CEH-generated images, these are currently stored on our in-house data centre. As our projects are taking place around the world, it is likely that our data will feed into national data centres once we have this all setup.
Does the image recognition only work with the setup you have described or could this be used on a range of light traps?
The model should be general enough to work across different setups. For example, in the Netherlands it has been used successfully with a different set-up that includes a yellow sheet.
Does the model only consider the image or can other factors (such as temperature, location etc.) be used to reach a species determination?
Currently, the only other factor considered by the model is location, so we have different models for different regions (for example, we have separate models for the UK, Costa Rica and northern USA/southern Canada). We are looking into other factors and did a test last year that investigated if moth activity and temperature correlated with each other. At the moment we will probably use things like phenology to check that the model matches what we already know. We already have a lot of abiotic data for the images gathered, such as date, vegetation around the AMI and sunset/sunrise times so there is potential to look at these factors in more detail in the future.