Our first 2025 BEGIN seminar – Dr. Weiming Huang
Date: Tuesday, February 25th, 2025
Time: Online
Geospatial Foundation Models
Slides
Abstract:
The rise of large-scale pre-trained models, also known as foundation models, has sparked great interest within the geospatial community. This trend naturally leads to two fundamental questions: 1) can general-purpose foundation models be directly applied to geospatial analyses? 2) given the transformative impact of large-scale pre-training in fields like natural language processing and computer vision, can we develop foundation models to benefit geospatial tasks?
In this presentation, I will share our recent explorations toward answering these questions. I will begin by exploring the application of various foundation models to a wide range of geospatial and urban analyses. In particular, I will present a study demonstrating how incorporating geospatial prior knowledge into the prompting process can dramatically improve urban land use inference using a vision-language model.
Taking a step further, we view learning effective representations from multi-modal geospatial data as the cornerstone of developing geospatial foundation models. To this end, I will present our progress in spatial representation learning across several widely used geospatial data sources, gradually leading to our recent work on developing geospatial foundation models, including an early version of City Foundation Models (CityFM).
Bio:
Weiming Huang is a Lecturer in Urban Data Science at the School of Geography, University of Leeds. He obtained his PhD in Geographical Information Science from Lund University, Sweden, and was a Wallenberg Postdoctoral Fellow at Nanyang Technological University, Singapore, and Lund University. He is the recipient of several prestigious awards, including the EuroSDR Award for the Best PhD Thesis Related to Geoinformation Science in 2021 and the Waldo Tobler Young Researcher Award from the Austrian Academy of Sciences in 2022. His research interests include spatial data mining, and more recently geospatial foundation models. He has served as a guest editor for IJGIS, TGIS, and JAG, and has been a program committee member for several top-tier machine learning conferences, such as ICLR, NeurIPS, and ICML.
