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

Abstract: Understanding urban spatial morphology is critical for climate modeling, risk assessment, and sustainable urban design, and Local Climate Zone (LCZ) mapping provides the basic framework for this. However, many cities still use coarse ~100-m resolution LCZ records, which are unsuitable for fine-scale urban research. In this study, precomputed embeddings from TESSERA (Feng et al., 2025) and AlphaEarth (Brown et al., 2025) are compared to traditional Sentinel-1/2 (S1S2) composites in five Swiss cities to see if they can upscale coarse LCZ maps to 10-m resolution using an attention-based U-Net. Three experiments assess multi-city transferability, the impact of higher-resolution reference data, and temporal robustness to year-to-year phenology changes. We find that all datasets achieve 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 outperforms both S1S2 and AlphaEarth across both settings As expected, we find that the transfer of embedding-based models from one year to another remains an open challenge. Overall, however, our results demonstrate the promising potential of embeddings derived from EO foundation models to reduce time consuming preprocessing, respectively, manual feature engineering tasks and to guide a universal deep learning-based LCZ mapping workflow. When combined with a simple location-aware attention U-Net architecture, the embeddings enhance regional transferability and scalability, supporting the development of comprehensive and reproducible fine-scale LCZ maps for global urban climate applications Improving reference data quality remains the strongest lever for further accuracy gains.

• The paper demonstrates that TESSERA embeddings achieve higher accuracy and spatial transferability (test IoU up to 0.82) than both AlphaEarth and Sentinel-1/2 composites for fine-scale LCZ mapping.
• The method leverages an attention-based U-Net with tailored input channels to integrate multimodal EO data, enhancing mapping performance for diverse urban morphologies.
• High-resolution labels and balanced sampling significantly boost performance, highlighting the potential for dynamic urban climate monitoring using EO embeddings.

Embeddings for Fine-Scale Local Climate Zone Mapping: Evaluation of AlphaEarth and TESSERA Across Swiss Cities

Motivation and Problem Statement

Local Climate Zones (LCZ) represent urban morphology more effectively than conventional Land Use Land Cover (LULC) classifications and thus are foundational for urban climate studies, informing risk assessments, climate modeling, and urban planning. However, the globally available LCZ datasets—most notably WUDAPT—remain limited to coarse resolutions (~100m), hindering their utility for applications requiring fine spatial granularity. The proliferation of Earth Observation (EO) foundation models, capable of generating dense pixel-wise embeddings (AlphaEarth, TESSERA), presents the opportunity to advance LCZ mapping via deep learning across diverse urban contexts at a higher spatial (10m) and temporal resolution. This paper conducts an empirical comparison of these EO embeddings with traditional Sentinel-1/2 composites for LCZ upscale in five major Swiss cities, deploying an attention-based U-Net architecture.

Dataset Description and Embedding Technologies

Three distinct data sources underpin the experiments:

• Sentinel-1/2 Seasonal Composites (S1S2): Baseline EO imagery covering the visible, near-infrared, and synthetic aperture radar bands, processed for cloud removal and atmospheric correction.
• AlphaEarth Embeddings: Google’s annual global embeddings at 10m derived from multimodal EO data (optical, radar, lidar, climate) using semi-supervised masked autoencoding. Embedding vectors are 64-dimensional, patch-based, and directly accessible via Google Earth Engine.
• TESSERA Embeddings: Provided at 10m via a fully self-supervised scheme leveraging Barlow Twins loss on temporal sequences of Sentinel-1 and Sentinel-2, producing 128-dimensional, pixel-level embeddings robust to missing observations and designed for gap-free spatial coverage.

Reference LCZ maps were sourced from WUDAPT (100m) for all cities, and an enhanced city-specific map for Bern at 78m enabled assessment of training reference resolution impact. The study focuses on Bern, Basel, Geneva, Lausanne, and Zurich, each representing varied urban morphologies and climates.

Deep Learning Architecture and Training Strategy

Each input dataset was mapped to LCZ classes using an attention-based U-Net, chosen for its modularity and spatial localization capacities. Separate channel configurations were defined for S1S2 (14/16 features), AlphaEarth (64), and TESSERA (128), backed by consistent training hyperparameters: Adam optimizer (lr=1e-3), batch size 16, Dice+Focal composite loss, and stratified patch sampling to mitigate class imbalance. Early stopping, learning rate scheduling, and model checkpointing were rigorously implemented to enforce generalization and prevent overfitting.

Three main experiment types were executed:

1. Multi-city spatial transfer: Model trained across all cities using coarse WUDAPT reference.
2. Single-city, high-resolution reference: Model trained on Bern using its higher-resolution LCZ map.
3. Temporal transferability: Models trained on 2024 data applied to predict 2025 LCZ maps, probing robustness to phenological shifts.

Numerical Results and Model Comparisons

Multi-City Mapping (Experiment I)

All datasets yielded robust segmentation, with TESSERA consistently outperforming S1S2 and AlphaEarth, achieving test IoU of 0.69 and overall accuracy of 0.80. AlphaEarth’s patch-based embeddings performed comparably to S1S2 but were marginally weaker in test generalization. Notably, all models demonstrated strong capacity to segment dominant urban and natural LCZ types (IoU ≥ 0.80 for classes like Compact Midrise, Large Lowrise, Dense Trees, Low Plants, Water). Underrepresented classes (e.g., Sparsely Built, Scattered Trees, Bush/Scrub) were poorly distinguished (IoU ≤ 0.70), reflecting persistent class imbalance and label ambiguity.

Single-City High-Resolution Reference (Experiment II)

Models trained with finer 78m labels in Bern achieved substantially higher test IoUs (up to 0.82 for TESSERA) and accuracy (up to 0.90). The increase in reference quality not only elevated absolute performance but also accentuated the advantage of embedding-based approaches in generalization, particularly for complex urban morphologies and rare LCZ types.

The S1S2 model, despite integration of high-resolution DSM/DTM topography, failed to outperform the embedding-based models in validation, suggesting embeddings’ superiority in capturing nuanced spectral-temporal signatures even in data-rich settings.

Temporal Transferability (Experiment III)

Direct cross-year application revealed stable spatial structure preservation across all approaches, but accuracy degraded modestly. S1S2 exhibited the highest temporal stability (accuracy declines ≤ 1%), whereas AlphaEarth suffered pronounced drops in some cities (e.g., Bern: 73.9% to 60.1%). TESSERA retained robust accuracy (e.g., test IoU decrease within 0.02–0.06 absolute units), attesting to the temporal resilience of self-supervised temporal embeddings, but also mirrored phenology-driven landscape changes. Both embedding providers commit to annual updates, supporting dynamic urban climate monitoring.

Cross-Reference Evaluation

Models trained on higher-resolution labels excelled when evaluated with similarly granular references (accuracy improvements up to 20 percent points), indicating that spatial resolution of labels is the primary lever for accuracy improvement. Conversely, coarse evaluation references bias down the assessment of genuinely fine-grained predictions, underlining the critical need for widespread high-resolution LCZ label acquisition.

Technical Implications and Performance Insights

• Embedding Advantages: TESSERA’s pixel-level temporal embeddings robustly encoded urban microstructure and phenology, yielding superior spatial transferability and generalization, especially in sparse and heterogeneous LCZ types. AlphaEarth’s patch-based embeddings were less effective in natural environments with high textural variability, but competitive for built-up urban zones.
• Attention U-Net Utility: The modular architecture supported adaptation to varying input feature dimensions and maintained robust segmentation quality across all cities, with minimal reliance on manual feature engineering or preprocessing.
• Resolution and Class Imbalance: Rare LCZ classes remained underserved, regardless of input modality, due to persistent imbalance and co-existence within pixels. Future research must prioritize balanced sample generation and the use of novel loss functions or embedding strategies to ameliorate this bottleneck.
• Temporal Resilience: Embedding models, particularly TESSERA, captured year-to-year urban landscape shifts, posing both opportunities for dynamic mapping and challenges for consistency in long-term climate studies.

Practical and Theoretical Implications

This research establishes the viability of embedding-driven LCZ mapping for scalable, reproducible urban climate zonation at 10m globally. The open-source nature of both inputs (embeddings, reference labels) and deep learning workflow democratizes access, lowering entry barriers for diverse stakeholders and facilitating continuous mapping as EO embeddings are updated.

From a theoretical vantage, the demonstrated spatial and temporal transferability of embedding representations signals a paradigm shift for geospatial segmentation tasks, moving away from reliance on hand-engineered features and manual preprocessing. The findings highlight the importance of embedding design—pixel-level vs. patch-level, temporal encoding, multi-sensor fusion—on downstream mapping performance and generalization.

Future developments may involve

• Integration of multi-year training for temporal robustness,
• Incorporation of additional urban or climatological features,
• Deployment of advanced architectures (e.g., transformers with spatial-temporal attention) for enhanced small-class detection,
• Expansion of cross-country and cross-continent benchmarks to verify universality.

Conclusion

The comparative evaluation underscores TESSERA’s superiority for fine-scale LCZ mapping across diverse urban morphologies, offering enhanced spatial transferability and accuracy relative to both patch-based AlphaEarth and traditional S1S2 composites. The greatest performance gains accrue from improvements in reference label resolution and balanced sample representation. Embedding-driven workflows present a transformative path for global, scalable, and reproducible LCZ mapping, setting the stage for widespread adoption in urban climate modeling, planning, and environmental monitoring. The technical results support continued investment in EO embedding development and high-resolution LCZ labeling initiatives, and the open-source dissemination of this methodology will spur further innovation and applications in geospatial AI.

---

Citation: "Exploring the potential of AlphaEarth and TESSERA embeddings for Fine-scale Local Climate Zone Mapping: A case study across five cities in Switzerland" (2606.20034)