LIDS Computing and Sustainability Seminar: Lauren Gillespie

"Opportunities and challenges for monitoring biodiversity in the Anthropocene with foundation models"

Climate and land use change have radically altered biodiversity worldwide. Foundation models trained on large-scale ecological and earth observation data are a promising new approach for monitoring these changes at global scale. This talk explores the opportunities foundation models offer for biodiversity monitoring, but also the challenges they face with inherently biased ecological data. First we’ll cover a new method for building biodiversity foundation models from citizen science observations and remote sensing data which significantly improves biodiversity detection and can estimate rapid change over space and time. Next, I’ll introduce a new ecologically-inspired machine learning technique for building biodiversity foundation models from heterogenous data without the need for extensive expert labeling. Last, we’ll explore how biases baked into many large-scale biodiversity datasets can influence foundation model behavior in real-world conservation settings. Looking forward, biodiversity foundation models have a role to play in monitoring the world’s progress towards the Global Biodiversity Framework’s 2030 Mission, but only if we acknowledge and adapt to their underlying limitations.