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

This study investigates whether precomputed geospatial embeddings—specifically TESSERA and AlphaEarth—can be used to generate fine-scale (10-meter) Local Climate Zone (LCZ) maps from coarse (100-meter) global reference data. The authors compare these foundation model embeddings against a baseline of traditional Sentinel-1 and Sentinel-2 (S1S2) seasonal composites. The researchers employed an attention-based U-Net architecture to perform this upscaling across five Swiss cities (Basel, Bern, Geneva, Lausanne, and Zurich), evaluating performance through multi-city transferability, the impact of reference data resolution, and temporal robustness.

The experiments demonstrate that using precomputed embeddings significantly simplifies the mapping workflow by eliminating the need for extensive manual feature engineering and preprocessing. All tested datasets achieved strong performance, with Intersection-over-Union (IoU) scores ranging from 0.59 to 0.69 in multi-city tests and 0.77 to 0.82 when using higher-resolution reference data. TESSERA consistently outperformed both AlphaEarth and the traditional S1S2 baseline. While the models showed promise for regional scalability, the authors emphasize that temporal transferability—maintaining accuracy when applying a model trained on one year to another—remains an open challenge.

Local Climate Zone mapping is essential for urban climate modeling and sustainable city planning, yet many regions rely on coarse 100-meter data that fails to capture fine-scale urban morphology. By leveraging ready-to-use embeddings from geospatial foundation models, researchers can reduce the technical and computational barriers to producing high-resolution, reproducible LCZ maps. This approach supports more accurate urban heat island assessments and climate risk modeling on a global scale, provided that high-quality reference data is available to guide the training process.