Mapping Local Climate Zones Globally with Geospatial Foundation Models

Speaker: Andrés Zúñiga González, University of Cambridge
Series: Institute of Computing for Climate Science (ICCS) Seminars

Local Climate Zones (LCZs) classify neighbourhoods by their physical form and land cover in a way that is comparable across cities, which makes them a key input for studying urban heat.

Existing global LCZ maps depend on volunteered training data of uneven quality and are costly to keep up to date. This is a particular problem for cities in the Global South, which are growing and changing rapidly, are often under-represented in training data, and are among the most exposed to extreme heat and other climate hazards.

Geospatial foundation models (GFMs) offer a cheaper route: they condense a year of satellite imagery into compact embeddings on which small, inexpensive models can be built. This project examines how well two such models, Cambridge's Tessera and Google's AlphaEarth, support global LCZ mapping.

Using expert-labelled LCZ datasets covering cities worldwide, a range of neural network architectures is trained to classify LCZs from these embeddings. The comparison shows how usable GFMs are for global downstream tasks and, more importantly, what it takes for LCZ models to transfer reliably to cities they have never seen.

Join on Zoom or in-person at MR13, Centre for Mathematical Sciences, Wilberforce Rd, Cambridge CB3 0WA