Exploring the potential of AlphaEarth and TESSERA embeddings for Fine-scale Local Climate Zone Mapping: A case study across five cities in Switzerland

2026-06-18 · Source: Machine Learning · Field: Science & Research — Environmental Science & Earth Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Precomputed embeddings from TESSERA (Feng et al., 2025) and AlphaEarth (Brown et al., 2025) are compared against traditional Sentinel-1/2 (S1S2) composites for upscaling coarse ~100-m resolution Local Climate Zone (LCZ) maps to 10-m resolution. This study, conducted across five Swiss cities, utilizes an attention-based U-Net architecture. Three experiments assessed multi-city transferability, the impact of higher-resolution reference data, and temporal robustness to phenology changes. All datasets achieved strong performance, with test data Intersection-over-Union (IoU) ranging from 0.59-0.69 and 0.77-0.82 in the first two experiments. TESSERA consistently outperformed both S1S2 and AlphaEarth. While temporal transferability remains a challenge, the results demonstrate the promising potential of embeddings from Earth Observation foundation models to reduce preprocessing and manual feature engineering, enhancing regional transferability and scalability for fine-scale LCZ mapping. Improving reference data quality is identified as the strongest lever for further accuracy gains.

Key takeaway

For AI Scientists developing urban climate models, you should consider integrating precomputed TESSERA embeddings with attention-based U-Net architectures for fine-scale Local Climate Zone mapping. This approach significantly reduces manual feature engineering and enhances regional transferability, supporting more reproducible and scalable global urban climate applications. Focus efforts on improving reference data quality, as this offers the most substantial accuracy gains, while acknowledging current challenges in temporal model transferability.

Key insights

Precomputed embeddings from EO foundation models can upscale coarse LCZ maps to 10-m resolution effectively.

Principles

• Embeddings reduce preprocessing and feature engineering.
• Reference data quality is key for accuracy gains.
• Temporal transferability of embedding-based models is challenging.

Method

An attention-based U-Net upscales coarse LCZ maps to 10-m resolution using precomputed TESSERA or AlphaEarth embeddings, or S1S2 composites.

In practice

• Use TESSERA embeddings for LCZ mapping.
• Prioritize high-quality reference data.
• Explore location-aware attention U-Net architectures.

Topics

• Local Climate Zone Mapping
• TESSERA Embeddings
• AlphaEarth Embeddings
• U-Net Architecture
• Urban Climate Modeling
• Remote Sensing

Best for: Computer Vision Engineer, AI Scientist, Research Scientist

Related on AIssential

• Physics-Conditioned Synthesis of Internal Ice-Layer Thickness for Incomplete Layer Traces — A novel network synthesizes complete ice-layer thickness from incomplete radar data by integrating geometric and transformer learning with physical climate model conditioning.
• Spatially-constrained clustering of geospatial features for heat vulnerability assessment of favelas in Rio de Janeiro — Settlement morphology significantly influences heat vulnerability in informal urban areas, with flat-terrain favelas experiencing higher temperatures.
• Task Aware Modulation Using Representation Learning for Upsaling of Terrestrial Carbon Fluxes — Integrating physical constraints with adaptive representation learning improves global carbon flux estimation accuracy.
• Characterizing AlphaEarth Embedding Geometry for Agentic Environmental Reasoning — AlphaEarth embeddings form a curved, heterogeneous manifold where retrieval is more reliable than vector arithmetic for environmental reasoning.
• AI Can Help Track the World’s Shrinking Glaciers — AI models for glacier tracking can achieve human-comparable accuracy in new regions with minimal targeted data and environmental context.
• Urban Vibrancy Embedding and Application on Traffic Prediction — Urban vibrancy embeddings derived from real-time floating population data significantly enhance traffic prediction accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.