From Lab to Field: AI for Automated Insect Monitoring
Guest: Yuyan Chen
Host: Cindy (Xinyi) He
Cohost: Hermione He
Research Analyst: Christy Zha, Raquel Chen, Ekki Lu
Meet Yuyan Chen, a first-year Master’s student in Computer Science at McGill University and researcher at Mila-Quebec AI Institute. With a passion for machine learning and computer vision, Yuyan channels her expertise toward addressing pressing global challenges, such as climate change and biodiversity conservation.
Currently immersed in research focused on insect monitoring and new species detection through camera-trap images, she is at the forefront of leveraging advanced AI techniques to revolutionize biodiversity monitoring.
But Yuyan’s journey into the realm of AI and environmental science began earlier, during her undergraduate years at McGill. After graduating from the Joint Honours Mathematics and Computer Science program, with a minor in Environment, she delved into multiple projects encompassing calcium diffusion in astrocytes, land cover type classification, topic modeling, and machine learning storage.
Beyond her academic pursuits, she is deeply committed to fostering equity, diversity, and inclusion (EDI) within STEM education. Yuyan is actively engaged in initiatives such as the Equity in Math Reading Club.
Join us as we delve into Yuyan Chen’s inspiring journey at the intersection of AI, environmental science, and social equity.
Can you provide an overview of the research conducted at Mila regarding Automated Monitoring of Insects using AI?
Mila is an AI research institute that consists of four to five universities in Montreal. Those labs are working on all kinds of machine learning in general. The Automated Monitoring Insects, or AMI, in particular, is an interdisciplinary project where we have several teams working on different parts. So for us, at Mila, we work on the software engineering and machine learning part.
The entire project starts from a camera trap, the hardware. Some teams are building hardware, and we also need to put the hardware in the wild, for example, in a rainforest. I had a chance to deploy camera traps in Panama last November. After we get videos of insects from the camera traps, we then use these data for machine learning. Our team at Mila fit in the last part of the pipeline: we build machine learning algorithms to detect insects, filter out non-moths, and identify their species with a fine-grained classifier.
My research focuses on using machine learning to discover the species that are not “seen” by the fine-grained classifier. These can be species that are new to science or well-known species that have only recently colonized the region. Software engineers on our team are also building a platform where people can just upload their data and run our model on their data.
Why do you focus on moths and what potential applications or benefits do you foresee from this research?
While our group primarily focuses on moths, a species often overlooked by the general public, we recognize their ecological significance. Despite the similarities between butterflies and moths, the latter are frequently disregarded due to their nocturnal habits and less vibrant appearances. However, moths play vital roles in ecosystems, serving as prey for birds, pollinators, and sometimes pests in agriculture.
With more than 160,000 species worldwide, far exceeding the number of butterfly species, it’s impractical for humans to identify them all manually. However, moths’ attraction to light makes them easier to monitor compared to butterflies. Hence, we aim to study moths systematically through an automated approach, acknowledging their importance despite their often underappreciated status.
Could you explain the specific AI techniques or algorithms being used in this research? How do they contribute to The Automated Monitoring Insects (AMI) processes? Any challenges?
Certainly. For us, the process is simplified due to our setup. We utilize camera traps with a white background, which attract moths to settle peacefully. In contrast, monitoring butterflies is more complex as it involves positioning the camera towards specific flowers that butterflies frequently visit.
In our approach, we employ computer vision techniques such as object detection to process the raw data captured by the camera traps, which often contain multiple moths within a single frame. Following detection, we utilize a binary classifier to filter out non-moth entities and then employ a fine-grained classifier to identify moth species. One notable challenge we face is that our training samples exhibit a long-tail distribution. This means that while we may have ample images for some classes, others have significantly fewer, resulting in an imbalanced dataset.
It’s important to emphasize that we do not trap or harm the moths in any way. We simply capture images, allowing the moths to continue their natural behavior unharmed. This stands in stark contrast to traditional trapping methods, which often result in the death of captured insects. Our approach yields a wealth of images alongside metadata, forming a comprehensive dataset for analysis.
How do you ensure the accuracy and reliability of the data collected through AI-based insect monitoring systems?
When I joined this project last May, my colleagues had already built an ML pipeline from object detection to fine-grained classification. The absence of labeled data was a significant challenge when my colleagues initially embarked on this project. Identifying moth species requires expertise typically possessed by entomologists or taxonomists, making it impractical to rely on random individuals for accurate labeling. Consequently, we turned to the rich dataset provided by GBIF, which comprises scientifically validated, high-resolution images. These served as our training data for the classifier.
However, a key discrepancy arose in the resolution between the high-quality images from GBIF and the lower-resolution images captured by our camera traps. To address this, we resize the images from the camera traps and use the approach of mixed-resolution augmentation during the training process, ensuring consistency in resolution. This approach mitigates domain adaptation challenges, where the model struggles to generalize from one domain (high-resolution images) to another (lower-resolution images).
What has been the most surprising or counterintuitive lesson you’ve learned through your work at the intersection of AI and environmental science?
I haven’t come across any counterintuitive findings in my research. However, when it comes to the broader application of AI, I believe it is crucial to exercise caution when employing AI for biodiversity monitoring or sustainability initiatives. Currently, large-scale models entail significant carbon emissions and consume substantial amounts of energy, comparable to that of a small country. This high energy consumption and carbon footprint are mainly incurred during inference rather than training, as these models are frequently utilized.
Therefore, when utilizing AI to address climate change or preserve the environment, it’s important to consider the energy consumption and carbon emissions associated with model usage. Failing to account for these factors could inadvertently exacerbate environmental harm. This underscores the importance of conscientiousness in AI research to ensure that our efforts align with sustainability goals rather than contradicting them.
Can you discuss any real-world applications or case studies where AI-powered insect monitoring has been implemented or could be implemented in the future?
I think it will be vision language models. This is how we combine computer vision with natural language such that we can actually generate descriptions that are helpful for experts’ work and science education.
What inspired you to pursue a career that combines computer science and environmental studies, and how has your background in both areas shaped your research approach?
In short, my interest in nature, environmental systems, and sustainability has been longstanding. Initially, during my undergraduate studies in chemical engineering, my focus was on addressing environmental issues like water pollution and climate change. However, my chemical engineering curriculum led me to explore computer science and mathematics, prompting a change in my major. It was during this transition that I came across a paper titled “Tackling Climate Change with Machine Learning”, which piqued my interest in the intersection of machine learning and sustainability. At the same time, I pursued a minor in environmental studies, creating a well-rounded foundation for my current work.
Before committing to computer science, I conducted extensive research online to understand how this field could contribute to environmental conservation. Discovering the paper I mentioned before, of which one of the authors later became my supervisor, played a pivotal role in solidifying my decision. While I appreciated the value of wet lab experiments, I ultimately gravitated towards computer science and mathematics due to my interest in coding and writing proofs and their potential to make meaningful contributions to addressing environmental challenges. The paper provided clarity and direction in my academic and professional journey.
Can you share with us your future plan?
Currently, I’m planning to pursue a PhD in this field because I really enjoy my research and think they are useful in terms of future career paths. I think anything related to sustainability would be really exciting for me. There are a lot of interesting AI applications in this domain. I recently learned about scalable fisheries monitoring with AI. Their algorithm automatically scans the number and the species of fish being brought onto a ship and helps report bycatch to facilitate sustainable fishery. There are also non-profits that are using satellite imagery and AI for marine debris detection and marine mammal monitoring.
What advice would you give to students who are interested in pursuing a similar path, combining AI and machine learning with environmental studies?
That’s a hard question because I didn’t know I would be working in this area when I started my undergrad. I just know that I would have worked out some things related to the environment. I think it is important to work on something that you find interesting and meaningful because research can be challenging and I get motivated by my interest. Also, keep your eyes open and learn about others’ research. And then you can find a lot of interesting things happening in this field.
Additionally, talk to a lot of people. Try to talk to people who are already working in the domain so you can get a sense of what your life will be like if you choose this path.
At the end of the interview, we discussed with Yuyan her experience studying/working at the fields of Entomology and Computer Science which are traditionally male-dominated.
As a feminist, Yuyan has read several articles discussing topics about women in academia and other fields. Here are two points she recognized during her journey: It is true that the representation of women in machine learning remains low. There are organizations like Women in Machine Learning that are working to encourage women to do ML research, but activities and organizations like that cannot fully solve the issue.
She thinks a good research and learning environment does not often remind her identity as a minority. On the one hand, it does create a place for minorities to speak freely, and everyone will have a sense of belonging, which she thinks is very good, but on the other hand, the phrase itself also emphasizes minorities being minorities:
“My advice to minorities in any field is to alleviate the burden of feeling obligated to prove the success of your minority group. This pressure often leads to increased stress and may exacerbate imposter syndrome. Instead, focus on your own growth and contributions, knowing that your efforts contribute positively to the diversity and richness of the field.”
Yuyan also recommended two articles discussing the possible imposter syndrome women often experience in this field (see references 1&2 below).
Read More:
The article that motivated her to change her major in CS: https://arxiv.org/abs/1906.05433
Project overview: https://globalnews.ca/news/10147726/montreal-research-artificial-intelligence-new-species/
Project proposal: https://www.climatechange.ai/papers/neurips2023/130
Yuyan’s Labmates’ paper: https://www.climatechange.ai/papers/neurips2023/47
Project field trip video: https://www.linkedin.com/feed/update/urn:li:activity:7162996166593363968/
