pith. sign in

arxiv: 2606.20034 · v1 · pith:NQEARCJCnew · submitted 2026-06-18 · 💻 cs.LG

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

Htet Yamin Ko Ko , Clement Atzberger This is my paper

Pith reviewed 2026-06-26 17:45 UTC · model grok-4.3

classification 💻 cs.LG
keywords Local Climate Zone mappingTESSERA embeddingsAlphaEarth embeddingsattention U-Neturban morphologytransferabilitySentinel-1/2 compositesfine-scale mapping
0
0 comments X

The pith

TESSERA embeddings outperform Sentinel-1/2 and AlphaEarth for upscaling coarse LCZ maps to 10m resolution with an attention U-Net across five Swiss cities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether precomputed embeddings from the TESSERA and AlphaEarth foundation models can replace traditional Sentinel-1/2 composites when upscaling existing coarse Local Climate Zone maps to 10-meter resolution. It runs three experiments with a location-aware attention U-Net on data from five Swiss cities to check how well models transfer between cities, how much better reference data helps, and whether results hold across different years. TESSERA embeddings deliver the highest accuracy in the multi-city and reference-data tests, with IoU scores of 0.59-0.69 and 0.77-0.82, while showing that foundation-model embeddings cut preprocessing and manual feature work. The study concludes that these embeddings support more universal and scalable LCZ workflows, though year-to-year transfer stays difficult and better reference data remains the main route to higher accuracy.

Core claim

Embeddings derived from EO foundation models reduce time consuming preprocessing and manual feature engineering tasks and guide a universal deep learning-based LCZ mapping workflow, enhancing regional transferability and scalability when combined with a location-aware attention U-Net. TESSERA embeddings consistently outperform both S1S2 and AlphaEarth across the first two experiments while year-to-year transfer remains an open challenge.

What carries the argument

Precomputed TESSERA and AlphaEarth embeddings used as input features to a location-aware attention U-Net for classifying Local Climate Zones at 10m resolution from coarse reference labels.

If this is right

  • Embeddings enable a single workflow for LCZ mapping that avoids city-by-city feature engineering.
  • Regional transferability improves, allowing one trained model to serve multiple cities.
  • Scalability for global urban climate applications increases due to lower preprocessing demands.
  • Further accuracy gains depend most directly on raising the quality of the reference LCZ data.
  • Temporal transfer across years requires separate solutions beyond the current embedding approach.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same embedding inputs could be tested on cities outside Switzerland to check whether the observed transferability extends globally.
  • Pairing these embeddings with other high-resolution reference sources might support even finer mapping scales.
  • The reduced preprocessing load could make detailed LCZ maps feasible for cities that currently lack them due to data or compute limits.
  • Finer LCZ layers produced this way would directly feed into higher-resolution urban climate and risk models.

Load-bearing premise

Coarse reference LCZ maps at roughly 100m resolution contain enough reliable information to train a model whose 10m outputs generalize across cities and are not driven mainly by city-specific phenology patterns.

What would settle it

If a model trained on four cities produces IoU below 0.5 on the held-out fifth city, or if accuracy falls sharply when the same model is applied to imagery from a different year with altered vegetation patterns.

read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper claims that precomputed embeddings from TESSERA and AlphaEarth foundation models, when paired with a location-aware attention U-Net, enable effective upscaling of coarse (~100m) LCZ reference maps to 10m resolution across five Swiss cities. It reports IoU ranges of 0.59-0.69 and 0.77-0.82 from three experiments testing multi-city transferability, higher-resolution reference impact, and temporal robustness to phenology changes, with TESSERA embeddings outperforming S1S2 baselines and AlphaEarth; the work concludes that such embeddings reduce preprocessing/feature engineering and support scalable, transferable LCZ workflows.

Significance. If the empirical results hold, the study provides concrete evidence that EO foundation model embeddings can simplify and improve regional transferability in LCZ mapping compared to traditional composites, with direct relevance to urban climate modeling and risk assessment. The multi-experiment design, explicit outperformance metrics, and acknowledgment of temporal transfer limitations constitute a proportionate contribution for an applied case study.

minor comments (3)
  1. Abstract: the reported IoU ranges are given without accompanying details on training/validation splits, hyperparameter choices, or exact experiment configurations; adding a brief methods summary would improve verifiability.
  2. The manuscript should clarify whether the attention U-Net architecture includes any city-specific adaptations or if it is strictly location-aware only via the stated mechanism.
  3. Consider adding a table summarizing per-city IoU scores and per-class performance to support the aggregate ranges and transferability claims.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work, the accurate summary of the manuscript, and the recommendation for minor revision. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper is a purely empirical case study that evaluates precomputed external embeddings (TESSERA, AlphaEarth) versus S1S2 baselines on held-out test IoU across three multi-city experiments; no equations, fitted parameters, or self-citations appear in the derivation chain, and the reported performance metrics are independent of any internal definitions or prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, invented entities, or ad-hoc axioms are stated in the provided text.

axioms (2)
  • domain assumption The LCZ classification scheme remains valid at 10 m resolution
    Implicit in the upscaling goal; standard assumption in the field.
  • domain assumption The attention U-Net architecture is appropriate for embedding-based semantic segmentation
    Chosen without justification in the abstract.

pith-pipeline@v0.9.1-grok · 5852 in / 1285 out tokens · 42039 ms · 2026-06-26T17:45:44.345304+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    G., Steiner, B., Tucker, P., Vasudevan, V ., Warden, P., … Zheng, X

    Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., Kudlur, M., Levenberg, J., Monga, R., Moore, S., Murray, D. G., Steiner, B., Tucker, P., Vasudevan, V ., Warden, P., … Zheng, X. (2016). TensorFlow: A system for large-scale machine learning. Ahamed, S., & Rahmim, A. (2023). Generalized Dice Foc...

    arXiv 2016

  2. [2]

    https://doi.org/10.1016/j.compag.2025.110676 Azad, R., Heidary, M., Yilmaz, K., Hüttemann, M., Karimijafarbigloo, S., Wu, Y ., Schmeink, A., & Merhof, D. (2023). Loss Functions in the Era of Semantic Segmentation: A Survey and Outlook. ArXiv. http://arxiv.org/abs/2312.05391 Bae, Y ., & Ban, Y . (2025). Class-Balanced Random Patch Training to Address Class...

    work page doi:10.1016/j.compag.2025.110676 2025

  3. [3]

    https://doi.org/10.3390/app15137056 Bhattiprolu, S. (2020). Smooth blending of patches for semantic segmentation of large images. Github. https://github.com/bnsreenu/python_for_microscopists/blob/master/229_smooth_predict ions_by_blending_patches/229_prediction_aerial_imagery_using_smooth_blending.py Brown, C. F., Kazmierski, M. R., Pasquarella, V . J., R...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.3390/app15137056 2020

  4. [4]

    Retrieved March 5, 2026, from https://developers.google.com/earth - engine/datasets/catalog/sentinel Guillaume Chevalier. (2017). Smoothly-Blend-Image-Patches. Github. https://github.com/V ooban/Smoothly-Blend-Image-Patches/blob/master/README.md Hamoudzadeh, A., Belloni, V ., & Ravanelli, R. (2026). Inferring Height from Earth Embeddings: First insights u...

    work page doi:10.1016/j.srs.2026.100397 2026

  5. [5]

    Le, Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N

    https://doi.org/10.3390/urbansci8030120 Oktay, O., Schlemper, J., Folgoc, L. Le, Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N. Y ., Kainz, B., Glocker, B., & Rueckert, D. (2018). Attention U-Net: Learning Where to Look for the Pancreas. Ramadhan Ramadhan. (2024). Land Cover Classification Using Deep Learning U -Net

    work page doi:10.3390/urbansci8030120 2018

  6. [6]

    (2026, March 4)

    https://github.com/ramiqcom/deep-learning-lc-unet University of Bern. (2026, March 4). UBELIX: High Performance Computing (HPC) of University of Bern. Github. https://www.id.unibe.ch/hpc Venkatachalam, C., Venkatachalam, S., & Balakrishnan, A. (2025). Enhanced skin cancer classification using modified efficientNetV2L with adaptive early stopping mechanism...

    work page doi:10.1038/s41598-025-22228-3 2026

  7. [7]

    S., Athanassiadis, A., & Binder, C

    https://doi.org/10.1016/j.cacint.2023.100135 Wiedmann, N. S., Athanassiadis, A., & Binder, C. R. (2023). Streamlining the regionalization of economy-wide material flow accounts (EW-MFA): The case of swiss cantons. Cleaner and Responsible Consumption, 10, 100127. https://doi.org/10.1016/j.clrc.2023.100127 Zbontar, J., Jing, L., Misra, I., Lecun, Y ., & Den...

    work page doi:10.1016/j.cacint.2023.100135 2023