PaperlayerView exact revision on arXiv

Source-linked AI summary

When local gains fail to transfer: Frozen Earth-observation embeddings across wildfires

Philipp Stark, Alexandros Sopasakis, Ola Hall

arXiv:2609.34602v1cs.LG

TL;DR

Frozen Earth-observation embeddings are commonly evaluated with spatially blocked validation within one region, while cross-regional and temporal generalization remain distinct evaluation questions. We assess them using pre-fire descriptors from wildfire events and a protocol combining held-out-fire transfer, local-label adaptation, and comparisons of fitted rules. The embeddings perform well within fires, but their advantage largely disappears across regions; nearby-fire forecasting in Attica provides a separate indication of cross-year transfer.

  • Problem

    How well frozen Earth-observation embeddings generalize across regions and time is less settled than their use as reusable features with task-specific classifiers.

    Source: §1 · p. 2

  • Method

    The study compares pre-fire TESSERA and AlphaEarth embeddings with WorldCover and Sentinel-2 descriptors across wildfire events, using held-out-fire transfer, limited local supervision, and fitted-weight comparisons.

    Source: §1 · p. 2, §7 · p. 10

  • Results

    Embeddings outperform conventional descriptors within fires, but their advantage largely disappears across regions; eight local blocks restore an advantage, while Parnitha-to-Varnavas TESSERA reaches 0.772 AUC, 0.040 below a within-Varnavas probe.

    Source: §5 · pp. 8–9

  • Takeaways & Limitations

    Frozen embeddings may be more useful as reusable representations for lightweight local models than as the basis for one transferable classifier.

    Source: §5 · p. 9

    The evaluation covers six fires, and temporal transfer rests on one pair, so it cannot establish how transfer varies with distance, landscape similarity, or years between fires.

    Source: §6 · p. 9

Abstract

from arXiv · show

Frozen Earth-observation embeddings are judged almost entirely by spatially blocked cross-validation inside one study region. We show that this number does not predict accuracy in a new region; we show why; and we show the one setting in which such a model does keep working, using a protocol that needs only a linear probe and labels one already has. The testbed is wildfire, with Copernicus burned-area maps of six fires in Greece and Spain and descriptors from the year before each fire, comparing TESSERA and AlphaEarth with ESA WorldCover classes and annual Sentinel-2 index summaries. Inside a fire, the embeddings identify the burned land 0.05 to 0.13 ROC AUC better than the index summaries, and repeated fold allocations, spatial buffers, a block bootstrap, and gradient-boosted trees leave that margin unchanged. On a fire in another region, they lose 0.15 to 0.18 AUC, and the index summaries lose 0.06, so the three end within a few hundredths of each other. The representation is not the cause. Eight labelled blocks from the new region restore the embedding advantage and give a higher AUC than 59,000 labelled pixels from other regions, and the weight vector fitted in one region is nearly orthogonal to the vector fitted in the others, so the part that carries across regions is small and low-dimensional. Forecasting within a region is a different matter. Fitted on a fire that burned in 2023 and applied to a fire twelve kilometres away that burned in 2024, where nothing used postdates the target fire, TESSERA reaches 0.772 AUC and loses 0.04 against a classifier fitted inside the 2024 fire, while classifiers fitted in other regions lose 0.09 to 0.18. A region with one mapped fire can therefore forecast susceptibility for later fires there; a region without one cannot borrow a model from elsewhere, and every evaluation of a frozen embedding should report a held-out region.

1 INTRODUCTION

This study asks whether frozen Earth-observation embeddings support classifiers that generalize across regions and years, and evaluates those dimensions using wildfire events. The embeddings outperform conventional descriptors within fires, but their cross-region advantage largely disappears.

1 INTRODUCTION

  • The study focuses on burned-versus-unburned pixels, comparing annual TESSERA and AlphaEarth embeddings with WorldCover classes and Sentinel-2 index summaries.It uses pre-fire descriptors and CEMS labels, and evaluates both cross-regional and temporal generalization.Source: §1 · p. 2
  • 0.05–0.13 AUC: frozen embeddings improve over Sentinel-2 index summaries within fires, with their combination performing best across all four fires.The combined embeddings reach ROC AUC 0.832–0.975.Source: §1 · p. 2
  • Across regions, the embedding advantage largely disappears, and a small amount of target-region supervision outperforms much larger training sets from elsewhere.This suggests the embeddings contain useful signal, but its relationship to wildfire occurrence is strongly region-specific.Source: §1 · p. 2
  • The Attica fires provide a within-region forecasting test: TESSERA retains much of its performance from the earlier to the later fire, while AlphaEarth has a weaker local advantage.Pooling the nearby fire with remote fires gives the best transfer for both embeddings; proximity and temporal closeness alone are not sufficient.Source: §1 · p. 2
  • The proposed protocol combines held-out-fire transfer, adaptation with limited local labels, and comparisons of fitted weight vectors.The study assembles six fires and about 120,000 labelled pixels across four descriptor sets, with fixed spatial fold assignments.Source: §1 · pp. 2–3

2 RELATED WORK

Prior work provides task-agnostic remote-sensing embeddings and benchmarks with frozen heads, while recent studies examine transfer in wildfire susceptibility and city mapping. This study focuses on event-level generalization across fires, regions, and time.

2 RELATED WORK

  • Earlier work provides task-agnostic remote-sensing embeddings and frozen-head benchmarks, including TESSERA and AlphaEarth as fixed global annual representations.The paper distinguishes its event-level focus from prior regional and city-level transfer studies.Source: §2 · p. 3
  • Prior studies report AlphaEarth transfer between nearby Australian regions for wildfire susceptibility and compare AlphaEarth and TESSERA across Swiss cities for local climate-zone mapping.The Swiss-city study examines both cross-city and temporal transfer.Source: §2 · p. 3
  • This evaluation treats each fire as an event, uses descriptors that predate it, and tests held-out fires with probes trained on different fires.Its sampling follows used–available designs, and its spatial validation follows established spatial-validation work.Source: §2 · p. 3

3 METHOD

The study evaluates pre-fire descriptors against mapped burned and control pixels from six fires, using spatially blocked probes and held-out-fire transfer. It also tests local-label adaptation, cross-fire learned rules, and forecasting between nearby fires in consecutive years.

3.1.1 WILDFIRE EVENTS

  • Four CEMS Rapid Mapping fires from 2023 form the main cross-regional set, with descriptors taken from 2022; the Attica pair adds Parnitha (2023) and Varnavas (2024).Parnitha and Varnavas are twelve kilometres apart, and each is represented by the year before it burned.Source: §3.1.1 · p. 3

3.1.2 BURNED AND CONTROL PIXEL SELECTION

  • Burned pixels are 10 m pixel centres inside mapped polygons after a 30 m interior buffer; controls are outside the polygons and a 500 m exclusion buffer.The sample comprises 5,000 burned and 15,000 control pixels per fire; controls represent locally available land, not what would have burned under other conditions.Source: §3.1.2 · pp. 3–4

3.1.3 FOUR DESCRIPTORS OF EACH PIXEL

  • The four descriptor sets are WorldCover classes, Sentinel-2 index summaries, TESSERA embeddings, and AlphaEarth embeddings.The index summaries use NDVI, NDMI, and NBR; TESSERA is dequantised using its per-pixel scale, and AlphaEarth is a separate, independently built embedding.Source: §3.1.3 · p. 4

3.2.1 PROBE AND WITHIN-FIRE EVALUATION

  • The primary probe is L2-regularised logistic regression, evaluated with ROC AUC and five-fold cross-validation holding out complete 5 km blocks.Because fold assignments matter with 39–90 blocks per fire, the study uses 21 spatial allocations and tests 500 m and 1 km buffers.Source: §3.2.1 · p. 4

3.2.2 ROBUSTNESS CHECKS

  • Robustness checks compare tuned logistic regression with LightGBM on persisted folds and test TESSERA dequantisation, nodata handling, and spatial buffers.Hyperparameters are selected within each outer training fold, not on its test fold.Source: §3.2.2 · p. 5

3.2.3 GENERALIZATION TO HELD-OUT FIRES

  • Held-out-fire transfer tests one-to-one and leave-one-fire-out classifiers, with paired fold comparisons and block-bootstrap uncertainty.Adaptation compares pooled remote-fire training plus 0–32 target blocks against training on those target blocks alone; a weight-vector analysis separates shared and target-specific directions.Source: §3.2.3 · p. 5
  • The temporal test trains on Parnitha and forecasts Varnavas using pre-fire descriptors and labels, then compares nearby-fire transfer with remote-fire sources and pooled combinations.The reverse transfer and target-label adaptation are also evaluated.Source: §3.2.3 · p. 5

4 RESULTS

Within fires, frozen embeddings outperform Sentinel-2 index summaries, but that margin largely disappears when probes transfer across regions. Local labels recover the advantage, consistent with region-specific fitted rules, while TESSERA transfers more effectively between nearby fires in the same region and successive years.

4.1 FROZEN EMBEDDINGS OUTPERFORM CONVENTIONAL DESCRIPTORS WITHIN FIRES

  • TESSERA exceeds Sentinel-2 summaries by +0.061 to +0.132 and AlphaEarth by +0.053 to +0.105 within fires; concatenating both embeddings gives the highest AUC on every fire.The paired embedding differences are positive in nearly every allocation, though neither embedding consistently leads the other.Source: §4.1 · pp. 5–6, Table 1 · p. 6

4.2 THE WITHIN-FIRE ADVANTAGE IS ROBUST

  • Across spatial buffers and block bootstrap, and with gradient-boosted trees, the within-fire embedding advantage remains; boosting preserves TESSERA’s lead over indices by +0.042 to +0.119.At a 1 km buffer, the TESSERA-minus-Sentinel-2 difference is +0.111 on Rhodes and +0.087 on Evros; dequantisation changes within-fire AUC by at most 0.006.Source: §4.2 · p. 6

4.3 THE EMBEDDING ADVANTAGE DISAPPEARS ACROSS REGIONS

  • Across regions, TESSERA loses 0.18 AUC on average and AlphaEarth 0.15, while Sentinel-2 indices lose 0.06; held-out mean AUCs are 0.69, 0.71, and 0.71, respectively.The sign of the held-out difference varies by fire; paired bootstrap intervals exclude zero only on Evros and Caceres.Source: §4.3 · pp. 6–7, Table 2 · p. 7

4.4 LIMITED LOCAL SUPERVISION RECOVERS THE EMBEDDING ADVANTAGE

  • With eight target-fire blocks, TESSERA beats indices trained on the same blocks by +0.12 on Rhodes and +0.04 on each other fire; adding 59,000 remote pixels makes TESSERA worse from eight blocks onward.With sixteen local blocks, TESSERA’s margins over indices are +0.13, +0.06, +0.09, and +0.05 across the four fires.Source: §4.4 · p. 7

4.5 THE LEARNED LINEAR RULE IS LARGELY REGION-SPECIFIC

  • Embedding probe weight vectors are nearly orthogonal across fires, with 91 to 99% of within-fire coefficient norm in the fire-specific remainder; the transferred component is low-dimensional and benefits from strong regularisation.For TESSERA, source-fire regularisation raises held-out mean AUC from 0.69 to 0.73, while within-fire AUC continues rising through 128 dimensions.Source: §4.5 · pp. 7–8

4.6 THE EMBEDDING ADVANTAGE IS RETAINED ONE YEAR LATER

  • From Parnitha to Varnavas, TESSERA reaches 0.772 AUC, 0.040 below the within-Varnavas result; combining Parnitha with pooled remote fires gives 0.783 for TESSERA and 0.761 for AlphaEarth.Reverse transfer is weaker: losses are 0.29 AUC for TESSERA and 0.16 for Sentinel-2 indices, so the Attica pair supports temporal transfer within one region, not a general embedding property.Source: §4.6 · p. 8

5 DISCUSSION

Frozen embeddings perform well within fires, but their advantage largely disappears across regions; limited local labels recover much of it. A nearby-fire test suggests TESSERA can also transfer across years, while the paper cautions that broader temporal stability remains unestablished.

5 DISCUSSION

  • TESSERA exceeds Sentinel-2 index summaries by 0.061–0.132 AUC within fires, but its advantage largely disappears across regions.Concatenating both embeddings gives the highest within-fire AUC for every fire, with a further 0.006–0.047 gain over the better single embedding; results remain stable across several evaluation choices.Source: §5 · p. 8
  • Eight labelled blocks from a target fire let TESSERA exceed same-data index summaries by 0.04–0.12 AUC, while fitted embedding weight vectors are nearly orthogonal across fires.Sixteen blocks recover much of the full within-fire advantage; adding approximately 59,000 pixels from other fires can reduce performance relative to target blocks alone.Source: §5 · p. 9
  • Frozen-embedding benchmarks with locally fitted heads measure useful information in one region, not whether the fitted rule generalizes elsewhere.The paper proposes reusable representations with lightweight local models, while noting that this reuse does not itself establish lower environmental cost.Source: §5 · p. 9
  • TESSERA reaches 0.772 AUC from Parnitha 2023 to Varnavas 2024, 0.040 below a probe trained within Varnavas.This is an initial indication from one Attica pair, not evidence of general temporal stability.Source: §5 · p. 9
  • Historical fires could supply local supervision for annual pre-fire descriptors, allowing region-specific models to rank susceptibility in later years.The authors describe monitoring or preventive-resource prioritisation as a potential application and say richer downstream models could be used without retraining the foundation model.Source: §5 · p. 9

6 LIMITATIONS

The evaluation covers six fires, leaving cross-regional comparisons with a small effective sample; temporal evidence is especially limited. Its environmental and baseline scope also constrains how broadly the observed results can be interpreted.

6 LIMITATIONS

  • Six fires leave cross-regional fire-level comparisons with a small effective sample; temporal analysis rests on one asymmetric pair.The study cannot establish how transfer changes with distance, landscape similarity, or years between fires; repeated fires across regions and years are needed before generalizing forecasting results.Source: §6 · p. 9
  • The fires cover Mediterranean and Macaronesian environments over two consecutive years, and the conventional baseline is intentionally simple.Stronger baselines using additional satellite bands, radar, terrain, weather, or other physical variables may reduce the observed embedding advantage.Source: §6 · p. 9
  • Different control-pixel composition across fires complicates interpretation of cross-regional loss, while the fire-specific embedding signal remains unexplained.Matched or land-cover-stratified controls and analysis linking the signal to land cover, vegetation, terrain, climate, and fuels are identified as next steps; operational use would require evaluation across substantially more regions and seasons.Source: §6 · p. 9

7 CONCLUSION

Across six wildfire events, frozen embeddings outperform conventional descriptors within regions, but that advantage largely disappears across regions. Limited local supervision recovers much of the advantage, and the temporal findings motivate broader evaluation of forecasting across years.

7 CONCLUSION

  • Within-region embedding performance exceeds conventional descriptors, but the advantage largely disappears under cross-regional generalization.Limited local supervision recovers much of the advantage, while the fitted linear rule is strongly region-specific.Source: §7 · p. 10
  • Temporal analysis suggests region-specific models can remain useful across years, motivating broader evaluation of temporal generalization for wildfire forecasting.The evaluation uses six wildfire events.Source: §7 · p. 10
  • The accompanying repository is described as containing analysis code, a labelled six-fire pixel table, fold assignments, probe probabilities, and an environment lock file.The authors also state that sampled AlphaEarth values are committed so the analysis does not require an Earth Engine account.Source: REPRODUCIBILITY STATEMENT · p. 10
  • The authors report using generative AI for language editing and some code improvements, with outputs reviewed, tested, and validated by the authors.They state that they take full responsibility for the manuscript and its contents.Source: AI USE STATEMENT · p. 10
  • The authors state that the study uses public satellite products and emergency-mapping polygons at 10 m resolution, with no personal data.The probes are research models, have not been validated for operational fire-risk assessment, and should not be used for it.Source: ETHICS STATEMENT · p. 10

A DATA AND PROCESSING DETAILS

The six-fire table distinguishes the four-fire benchmark from the two-fire Attica pair and records pre-fire data coverage. Inputs include fixed TESSERA and AlphaEarth embeddings and Sentinel-2 observations.

A DATA AND PROCESSING DETAILS

  • Table 4 covers six wildfire events: four benchmark fires and the two-fire Attica pair, with 5 km spatial blocks and pre-fire observation counts.No-data pixels are control pixels with no TESSERA embedding and are excluded from all analyses.Source: Table 4 · p. 13
  • Sentinel-2 inputs are pre-fire-year Level-2A scenes, with B04 and B08 sampled at 10 m and B11 and B12 bilinearly resampled from 20 m.Scenes came from the Planetary Computer STAC; pixels without valid observations were dropped.Source: A DATA AND PROCESSING DETAILS · p. 13
  • TESSERA uses v1.0 variant vultr embeddings from the pre-fire year, while AlphaEarth uses annual A00–A63 vectors sampled at 10 m.TESSERA embeddings are stored as int8 values with a per-pixel scale; sampled AlphaEarth vectors have unit norm.Source: A DATA AND PROCESSING DETAILS · p. 14

B ADDITIONAL WITHIN-FIRE RESULTS

Within-fire results remain favorable to the embeddings under alternative classifiers and spatial checks, while the fire-specific landscape contrast and low-dimensional analyses qualify what those results represent.

B ADDITIONAL WITHIN-FIRE RESULTS

  • Figure 3 shows sampled burned and control pixels, WorldCover classes, and 3-D TESSERA visualisations for Rhodes and Evros.Source: Figure 3 · p. 14
  • Figure 4 compares within-fire AUC across 21 block-to-fold allocations for the two Greek fires, marking individual allocations, their mean, and the default split.Source: Figure 4 · p. 15
  • Boosting lifts index summaries by +0.005 to +0.036, TESSERA by +0.003 to +0.033, and AlphaEarth by +0.008 to +0.031.On held-out fires, boosted TESSERA AUCs are 0.63, 0.65, 0.69, and 0.86, against 0.60, 0.70, 0.65, and 0.85 for the indices.Source: B ADDITIONAL WITHIN-FIRE RESULTS · p. 15, Table 5 · p. 15
  • A 1 km training-to-test buffer lowers TESSERA AUC from 0.846 to 0.820 on Rhodes and from 0.855 to 0.839 on Evros, while Sentinel-2 reaches 0.709 and 0.753.The buffer removes about 14% of training pixels.Source: B ADDITIONAL WITHIN-FIRE RESULTS · p. 15
  • Tenerife reaches 0.81 AUC with land-cover class alone; its burned–control contrast is largely a land-cover contrast because burns are mostly pine forest.Tenerife is also the easiest target for probes trained on Evros and the worst source for the other fires.Source: B ADDITIONAL WITHIN-FIRE RESULTS · p. 15
  • Table 6 describes a decomposition of each within-fire probe direction into the direction learned from the other fires and a fire-specific remainder.It compares test-fold AUC using the shared direction, the remainder, and the full within-fire direction.Source: Table 6 · p. 16
  • Within-fire embedding AUC continues rising through full PCA rank, whereas held-out AUC peaks at a few components.Source: Figure 5 · p. 16

C FULL TRANSFER RESULTS

Remote-fire transfer remains variable across source–target pairs, and boosting does not restore the within-fire margin. The nearby Attica pair provides a more favorable, but direction-dependent, temporal transfer case.

C FULL TRANSFER RESULTS

  • Five of twelve pairwise directions favor the indices over TESSERA and seven favor TESSERA, while the indices beat AlphaEarth in eight; pairwise and leave-one-fire-out designs agree.Increasing training fires from one to three raises mean held-out AUC by about 0.01 for TESSERA and 0.08 for AlphaEarth, without restoring the within-fire margin.Source: C FULL TRANSFER RESULTS · p. 16
  • Boosted held-out TESSERA AUCs are 0.63, 0.65, 0.69, and 0.86, versus 0.60, 0.70, 0.65, and 0.85 for the indices; boosted AlphaEarth falls below its linear probe on every fire.Table 8 reports leave-one-fire-out transfer ROC AUC with boosting tuned inside the training fires.Source: C FULL TRANSFER RESULTS · p. 16, Table 8 · p. 17
  • From Parnitha to Varnavas, whole-fire AUC is 0.738 for TESSERA, 0.719 for AlphaEarth, and 0.669 for indices; reverse transfer gives 0.619, 0.695, and 0.593.Embeddings lead the indices in both directions, but only the forward direction retains most within-fire performance.Source: C FULL TRANSFER RESULTS · pp. 16–17
  • Table 7 presents pairwise and leave-one-fire-out whole-target-fire logistic-transfer AUCs, excluding nodata; whole-fire rankings differ from fold-averaged Table 2 values.Source: Table 7 · p. 17

D TWO PITFALLS WE MET - BEYOND THIS STUDY

The authors identify two pitfalls: repeated grouped-fold splits can conceal split variability, and zero-vector missing embeddings can act as a label shortcut.

D TWO PITFALLS WE MET - BEYOND THIS STUDY

  • For Rhodes, 21 StratifiedGroupKFold seeds produced only three distinct splits, so seed-based standard deviation did not measure split-to-split variability.The authors instead drew and saved 21 random block-to-fold allocations for all experiments.Source: D TWO PITFALLS WE MET - BEYOND THIS… · p. 17
  • All-zero TESSERA vectors, nearly all permanent water controls, can let classifiers learn a control shortcut; the observed effect was at most 0.004 AUC.The effect grows with class imbalance, and cross-validation retaining these pixels cannot detect it.Source: D TWO PITFALLS WE MET - BEYOND THIS… · p. 17

E DESCRIPTIVE RESULTS FOR THE GREEK FIRES

For the Greek fires, burned and control pixels are separated in embedding space overall and within qualifying land-cover classes. Burned pixels are also over-represented in tree cover and shrubland.

Loading 2609.34602v1…