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

Philipp Stark; Alexandros Sopasakis; Ola Hall

cs.LGSep 28, 2026

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

Authors: Philipp Stark, Alexandros Sopasakis, Ola Hall

Organizations: Department of Human Geography Lund University Sölvegatan 10, SE-223 62 Lund, Sweden · Centre for Mathematical Sciences Lund University Sölvegatan 18, SE-221 00 Lund, Sweden

Abstract

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.

Figures & tables

Representation	Rhodes $\uparrow$	Evros $\uparrow$	Cáceres $\uparrow$	Tenerife $\uparrow$
WorldCover	$0.629\pm 0.014$	$0.639\pm 0.011$	$0.597\pm 0.028$	$0.814\pm 0.038$
Sentinel-2 (15)	$0.693\pm 0.028$	$0.745\pm 0.009$	$0.763\pm 0.044$	$0.890\pm 0.030$
Sentinel-2, no counts	$0.681\pm 0.019$	$0.742\pm 0.011$	$0.750\pm 0.047$	$0.887\pm 0.042$
TESSERA	$0.825\pm 0.026$	$0.853\pm 0.007$	$0.824\pm 0.061$	$0.958\pm 0.014$
AlphaEarth	$0.746\pm 0.033$	$0.850\pm 0.008$	$0.851\pm 0.035$	$0.969\pm 0.009$
AlphaEarth + TESSERA	$\mathbf{0.832}\pm 0.027$	$\mathbf{0.900}\pm 0.006$	$\mathbf{0.861}\pm 0.032$	$\mathbf{0.975}\pm 0.009$

Table 1: Within-fire ROC AUC, logistic regression, mean

\pm

SD over 21 block-to-fold allocations, 79,056 pixels. TESSERA is the dequantised embedding. The lower rows are paired differences computed within fold, with the number of allocations in which the difference is positive. Bold: best representation per fire. Arrows give the direction of improvement.

	TESSERA			AlphaEarth			S2 indices
Held-out fire	within $\uparrow$	held out $\uparrow$	loss $\downarrow$	within $\uparrow$	held out $\uparrow$	loss $\downarrow$	within $\uparrow$	held out $\uparrow$	loss $\downarrow$
Rhodes	0.825	0.595	0.230	0.746	0.606	0.140	0.681	0.632	0.049
Evros	0.853	0.674	0.179	0.850	0.681	0.170	0.742	0.718	0.024
Cáceres	0.824	0.697	0.128	0.851	0.746	0.105	0.750	0.662	0.089
Tenerife	0.958	0.793	0.165	0.969	0.793	0.177	0.887	0.831	0.056
mean	0.865	0.690	0.176	0.854	0.707	0.148	0.765	0.711	0.055

Table 2: Fire-level summary on matched spatial folds. Within: probe trained inside the fire. Held out: probe trained on the other three fires, scored on the same folds of the same 21 allocations. Loss: within minus held out, paired within fold. S2 = Sentinel-2 index summaries without observation counts. Logistic regression is used here. TESSERA is dequantised. Bold: best of the three representations per fire and column. Arrows give the direction of improvement.

Figure 1: Adaptation curves. AUC on the held-out target fold against the number of labelled target blocks in training, for probes trained on those blocks alone (solid) and on the three other fires plus those blocks (dashed). Mean over five folds and two allocations. The point at zero is the leave-one-fire-out probe.

	TESSERA		AlphaEarth		S2 indices
Training source	AUC $\uparrow$	loss $\downarrow$	AUC $\uparrow$	loss $\downarrow$	AUC $\uparrow$	loss $\downarrow$
Rhodes, 2023	0.675	0.137	0.616	0.227	0.757	0.057
Evros, 2023	0.628	0.183	0.697	0.146	0.790	0.025
Cáceres, 2023	0.629	0.183	0.672	0.171	0.638	0.176
Tenerife, 2023	0.721	0.091	0.758	0.085	0.755	0.059
Parnitha, 12 km away, 2023	0.772	0.040	0.740	0.103	0.649	0.165

Table 3: Transfer to the Varnavas 2024 fire from different training sources, evaluated on Varnavas’s spatial folds (allocation 0, mean of five folds). Loss is the AUC difference relative to a probe trained within Varnavas. Bold indicates the best training source for each descriptor.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Activation	Fire	Date	CEMS product	Burned (ha)	Blocks	Nodata	S2 obs.
EMSR675	Rhodes, Greece	Jul 2023	DEL_MONIT08	17,770	48	440	122
EMSR686	Evros, Greece	Aug 2023	DEL_MONIT09	93,939	90	338	122
EMSR667	Cáceres, Spain	May 2023	DEL_PRODUCT v2	9,595	47	0	228
EMSR685	Tenerife, Spain	Aug 2023	DEL_MONIT02	10,960	39	162	58
EMSR690	Parnitha, Greece	Aug 2023	DEL_MONIT03	6,190	29	0	148
EMSR746	Varnavas, Greece	Aug 2024	DEL_MONIT02	10,405	20	0	128

Appendix

Table 4: The six wildfire events. The first four are the benchmark. The last two are the Attica pair of Section 4.6 , with descriptors from 2022 and 2023 respectively. Blocks are 5 km spatial cross-validation blocks. No-data pixels are control pixels with no TESSERA embedding (all-zero vector), nearly all permanent water. They are excluded from all analyses. Sentinel-2 observations are the median number of cloud-free pre-fire observations per pixel.

Figure 2: AUC on each fire’s own spatial folds when the probe is trained within the fire (filled) and when it is trained on the other three fires (open), for TESSERA, AlphaEarth, and the Sentinel-2 index summaries.

Figure 3: Sampled burned and control pixels, WorldCover class, and a three-dimensional TESSERA visualisation for Rhodes (top) and Evros (bottom).

Figure 4: Within-fire AUC under each of 21 block-to-fold allocations for the two Greek fires. Grey points are single allocations, bars are their mean, and diamonds are allocation 0, the one produced by scikit-learn’s default splitter.

Fire	LR $\uparrow$	GBM $\uparrow$	LR $\uparrow$	GBM $\uparrow$	LR $\uparrow$	GBM $\uparrow$	LR $\uparrow$	GBM $\uparrow$	LR	GBM
	Sentinel-2 (15)		S2 no counts		TESSERA		AlphaEarth		$\Delta$ (T $-$ S2)
Rhodes	0.700	0.705	0.691	0.706	0.821	0.824	0.734	0.743	$+0.120$	$+0.119$
Evros	0.748	0.784	0.748	0.759	0.848	0.858	0.846	0.877	$+0.100$	$+0.074$
Cáceres	0.774	0.791	0.758	0.767	0.823	0.856	0.866	0.894	$+0.048$	$+0.064$
Tenerife	0.891	0.926	0.885	0.923	0.955	0.968	0.968	0.976	$+0.065$	$+0.042$

Appendix

Table 5: Within-fire ROC AUC with a tuned logistic regression (LR) and with gradient boosting (GBM) on the same persisted folds (allocations 0 to 4, mean over folds and allocations, TESSERA raw integers, which differ from the dequantised embedding by at most 0.006 within fire). The last two columns are the paired TESSERA minus Sentinel-2 (15) difference under each classifier. Bold: best representation per fire under each classifier.

	TESSERA				AlphaEarth				S2 indices
Fire	cos	shared	specific	full	cos	shared	specific	full	cos	shared	specific	full
Rhodes	0.12	0.594	0.837	0.843	0.20	0.605	0.770	0.781	0.54	0.634	0.625	0.697
Evros	0.20	0.674	0.847	0.859	0.41	0.678	0.747	0.843	0.24	0.719	0.746	0.751
Cáceres	0.22	0.699	0.859	0.869	0.03	0.752	0.868	0.870	$-0.25$	0.676	0.765	0.767
Tenerife	0.16	0.817	0.955	0.957	0.09	0.793	0.957	0.959	0.71	0.855	0.831	0.912

Appendix

Table 6: Within-fire probe direction decomposed along the direction learned on the other three fires (allocation 0, mean of five folds, global standardisation). cos: cosine between the two directions. AUC on the test fold using the shared direction alone, the fire-specific remainder alone, and the full within-fire direction. Bold: larger of shared and specific.

Figure 5: AUC of the probe on the top

k

principal components, within-fire (PCA fitted on the training folds) and held-out (PCA fitted on the three training fires), allocation 0. Within-fire AUC of the embeddings keeps rising to full rank. Held-out AUC peaks at a few components.

Train	Test	S2 (15) $\uparrow$	S2 idx $\uparrow$	TESSERA $\uparrow$	T. deq. $\uparrow$	AE $\uparrow$	AE+T $\uparrow$
Rhodes	Evros	0.640	0.663	0.654	0.642	0.567	0.639
Rhodes	Cáceres	0.526	0.604	0.620	0.617	0.684	0.732
Rhodes	Tenerife	0.847	0.828	0.777	0.786	0.589	0.678
Evros	Rhodes	0.562	0.632	0.568	0.563	0.553	0.495
Evros	Cáceres	0.717	0.702	0.676	0.654	0.663	0.684
Evros	Tenerife	0.847	0.849	0.753	0.766	0.826	0.873

Appendix

Table 7: Pairwise and leave-one-fire-out transfer ROC AUC over the whole target fire, logistic regression, nodata excluded. Whole-fire AUCs differ from the fold-averaged values of Table 2 because they also rank pixels across folds. Bold: best representation per row. S2 (15) includes observation counts and S2 idx excludes them. T. deq. = TESSERA dequantised, AE = AlphaEarth, AE+T = AlphaEarth and TESSERA concatenated.

Held-out fire	S2 (15) $\uparrow$	S2 idx $\uparrow$	TESSERA $\uparrow$	TESSERA deq. $\uparrow$	AlphaEarth $\uparrow$
Rhodes	0.645	0.601	0.630	0.632	0.550
Evros	0.638	0.697	0.651	0.668	0.658
Cáceres	0.342	0.648	0.687	0.681	0.680
Tenerife	0.798	0.852	0.859	0.861	0.806

Appendix

Table 8: Leave-one-fire-out transfer ROC AUC with gradient boosting, tuned inside the training fires. Bold: best representation per row.

Explore similar work

Aug 12, 2026stat.AP

Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping

Wildfire susceptibility mapping typically relies on physical variables assembled from multiple remote-sensing, climate, and geospatial products. AlphaEarth Foundations (AEF) provides analysis-ready geospatial embeddings that may reduce this dependence on heavy harmonisation and task-specific feature engineering, but their value for wildfire susceptibility mapping has not been systematically evaluated. Using Victoria, Australia (2017-2025), as a case study, we show that AEF embeddings can reconstruct commonly used variables in wildfire susceptibility analysis with high accuracy. In downstream susceptibility models trained on satellite-derived fire occurrence data, embedding-based susceptibility models achieve ROC-AUC values above 0.92 and consistently identify high wildfire susceptibility across eastern Victoria, particularly Gippsland and the north-eastern uplands, with additional localized hotspots in central and northwestern Victoria. A key feature of AEF embeddings is their strong near-region transferability within climatically similar regions. When embedding-based models trained in Victoria are applied to Canberra and Western Sydney-Blue Mountains, ROC-AUC improves by around 4% at Canberra and declines by around 2% at Western Sydney-Blue Mountains, compared with a mean decrease of approximately 25% for physical-variable models. These findings provide practical guidance for using AEF embeddings and lay a foundation for scalable wildfire susceptibility mapping workflows for downstream users such as government agencies and (re)insurers.

Yuan Zhuang, Sanaa Hobeichi, Peng Shi +1

1UNSW Sydney, School of Risk and Actuarial Studies, New South Wales, Australia · 2ARC Centre of Excellence for the Weather of the 21st Century, Australia · 3Climate Change Research Centre, University of New South Wales, Sydney, New South Wales, Australia +2

Wildfire Prediction Geospatial Foundation Models

Ask PDF

May 14, 2026cs.LG

Does Your Wildfire Prediction Model Actually Work, or Just Score Well?

Wildfire prediction is important for early warning and resource allocation, yet existing Earth foundation models (Earth FMs) are pretrained for general atmospheric and geophysical objectives rather than wildfire forecasting. To address this gap, we introduce WILDFIRE-FM, the first foundation model pretrained specifically for wildfire prediction using weather, active-fire observations, topography, vegetation, and static environmental data. However, introducing a domain-specific backbone alone does not solve the evaluation problem: wildfire events are sparse in space and time, making transfer conclusions highly sensitive to matching rules and evaluation settings. To address this problem, we introduce a fixed-contract evaluation framework with two controlled checks: a fixed-output check for matching-rule effects and a fixed-feature check for head-selection effects. Under matched contracts, we compare WILDFIRE-FM with ten Earth-FM baselines across occupancy, spread, retrieval, and regression tasks. Our results show that wildfire transfer conclusions depend strongly on evaluation design and task formulation. We hope this framework and WILDFIRE-FM provide a foundation for future wildfire-specific Earth-FM research and benchmarking. Our code is available at https://anonymous.4open.science/r/Wildfire-fm-evaluation-contracts-5AE9/.

Yangshuang Xu, Yuyang Dai, Liling Chang +2

Florida State University · Northeastern University

Wildfire Prediction Wireless Foundation Models Foundation Model Scores

Ask PDF

May 18, 2026cs.CV

Better Together: Evaluating the Complementarity of Earth Embedding Models

Earth embedding models transform Earth observation data into embeddings uniquely tied to locations on the Earth's surface. These models are typically evaluated in isolation, comparing the downstream task performance across different Earth embeddings. However, spatially aligned embeddings can naturally be fused, providing richer information per location, a capability that isolated evaluations fail to capture. We therefore propose assessing Earth embeddings by their complementarity: the performance gain of fused embeddings over the best single-model baseline. To operationalise this, we introduce an embedding complementarity index applicable to any embedding and task, and evaluate four Earth embedding models (AlphaEarth, Tessera, GeoCLIP, SatCLIP) in isolation, in all pairs, and jointly across six downstream tasks. Fused embeddings outperform the best single model in four out of six tasks, confirming that single-embedding evaluations often underestimate Earth embedding capabilities. Complementarity proves both task- and location-dependent. Further, for a land cover regression task, we find that complementarity is partially determined by the spatial scale of land cover classes. Complementarity reframes Earth embeddings: the greatest future gains may come not from any single Earth embedding model, but from combinations that are better together.

Thijs L van der Plas, Jacob JW Bakermans, Vishal Nedungadi +3

Wageningen University, NL · University College London, UK · University of Bonn, Germany

Earth Observation Land Cover Earth Complementarity

Ask PDF