Abstract

Abstract

While the growing availability of satellite data has revolutionized large-scale vegetation monitoring, existing global habitat maps remain coarse and often inaccurate, leaving us without a reliable system for tracking habitat change. Current disturbance maps typically oversimplify complex dynamics into binary categories (e.g., forest vs. non-forest) while static habitat maps rely on rigid thresholds and abrupt transitions that mask natural ecological gradients. We propose that Geospatial Foundation Model (GFM) embeddings, such as those from Alpha Earth and Tessera, provide a unified physical state space capable of representing both ecological continuity and dynamic transitions. In this talk, I will present early evidence suggesting that the latent geometry of these models can capture continuous spatial gradients, as well as represent ecological changes as coherent directional trajectories within the embedding space. These findings suggest that GFMs may offer a continuous, data-driven substrate for ecology, enabling a shift from static classification to a dynamic state-transition model that can quantify both the precise state of an ecosystem and its trajectory of change.

Bio

Jovana Knezevic is a second-year PhD researcher in the Department of Plant Sciences at the University of Cambridge. Leveraging a background in Computer Science and a decade of experience in the tech industry, her research focuses on bridging the gap between geospatial foundation models and applied ecology.