Annual Earth-observation embeddings encode wildfire disturbance and support simplified burned area mapping

Jovana Knezevic; Clement Atzberger; Zhengpeng Feng; Adam F. A. Pellegrini; Srinivasan Keshav; David Coomes

Figure 1: Distribution and composition of the HLS Burn Scars benchmark across the contiguous United States (2018–2021). (a) Chip locations by biome; (b) chips by fire year; (c) fires by incident type; (d) burned and unburned pixels; (e) MTBS burn severity; and (f) pre-fire land cover. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Analysis	Cohort and evaluation
Latent-space analysis (Section 3.1 )	1,282 eligible fire–chip pairs from 793 chips; balanced burned and unburned samples from refined MTBS masks across $E_{-1}$ , $E_{0}$ , and $E_{+1}$ ; separability scored per pair
Single-fire mapping (Section 3.2.2 )	Single-fire subset: 559 chips; refined MTBS masks; three-fold spatial CV; full-chip, pixel-pooled scoring
Annual BA mapping (Section 3.2.3 )	793 chips; union of all same-year refined MTBS burned areas; three-fold spatial CV; full-chip, pixel-pooled scoring
California wall-to-wall (Section 3.3 )	Train: 696 non-California chips with refined MTBS masks. Evaluate: statewide 2021 reference from 37 MTBS fires and 386 CAL FIRE perimeters
European transfer (Section 3.4 )	Train: 793 US chips from 2018–2021 with refined annual-fire MTBS labels. Evaluate: 88 European EMSR fire AOIs from 2024–2025, one per fire
Fire timing (Section 3.5 )	Train: burned pixels from 793 chips using MTBS ignition DOY. Evaluate US: 1,300 eligible fire–chip pairs using three-fold spatial CV; pixel predictions averaged for each pair. Evaluate Europe: 88 EMSR fires; pixel predictions averaged per fire and compared with the earliest post-event image date.

Table 1: Primary experimental cohorts and evaluation units. Unit definitions are provided in Section 2.2 .

Model	Type	Key configuration	Training configuration
Logistic regression	Linear pixel classifier	L2 regularization; $C=1$ ; balanced class weights; standardized inputs	—
Random Forest	Non-linear pixel classifier	200 trees; $\sqrt{d}$ features/split; unrestricted depth; min. leaf 5; min. split 10	—
Standard U-Net	Spatial segmentation	ResNet-18 encoder, randomly initialized; $\sim$ 14.5 M parameters	BCE + Dice; AdamW ( $lr=10^{-4}$ ); batch 32; max. 50 epochs; patience 10
Lightweight U-Net	Spatial segmentation	3-level encoder–decoder; base width 32; $\sim$ 0.87 M parameters	BCE + Dice; AdamW ( $lr=10^{-4}$ ); batch 64; max. 50 epochs; patience 10

Table 2: Downstream models used in burned area mapping experiments.

Figure 2: Spatial organization of pre-fire and fire year embeddings for an illustrative fire–chip pair. Left panels show pre-fire land cover and the reference burned area. (a–b) RGB composites of the first three principal components of Tessera and AlphaEarth embeddings for

E_{-1}

and

E_{0}

; PCA was fitted jointly across both years for each embedding product. (c–d) Spatial assignments from unsupervised k-means clustering (

k=2

), fitted separately to each embedding year.

Figure 3: Wildfire disturbance reorganizes annual Earth-observation embedding space. (a–d) PCA projections for an illustrative fire–chip pair, with PCA fitted jointly to (

E_{-1}

) and (

E_{0}

) embeddings of each product. Points are colored by pre-fire land cover (a–b), or burn status in fire year (c–d). (e–f) Separability across all eligible fire–chip pairs: (e) agreement between k-means clusters and burn labels, where 0.5 indicates chance agreement; and (f) cosine distance between burned and unburned centroids. Lines show medians and shaded regions the interquartile range.

Figure 4: Burned area delineation across input representations and downstream models on the 559 single-fire chips. Bars show mean burned-class (a) F1 and (b) IoU across three spatially blocked cross-validation folds, and error bars show

\pm

1 standard deviation. Full metrics are provided in Table S1 of the Supplementary Material.

Figure 5: Qualitative comparison of single-fire burned area delineation across input representations. Columns show (a) benchmark post-fire HLS image as a SWIR2–NIR–Red false-colour composite, (b) the refined MTBS reference, and predictions from (c) Tessera, (d) AlphaEarth, (e) post-fire HLS, and (f) paired pre/post HLS. Black indicates correct burn detection, blue omission error, and orange commission error. Rows were selected to illustrate characteristic performance regimes across approaches.

Representation	Model	IoU $\uparrow$	F1 $\uparrow$
Tessera	LogReg	$0.642\pm 0.041$	$0.781\pm 0.031$
Tessera	RF	$0.739\pm 0.038$	$0.849\pm 0.025$
Tessera	Light U-Net	$\mathbf{0.813\pm 0.027}$	$\mathbf{0.896\pm 0.016}$
AlphaEarth	LogReg	$0.241\pm 0.033$	$0.388\pm 0.043$
AlphaEarth	RF	$0.402\pm 0.053$	$0.571\pm 0.054$
AlphaEarth	Light U-Net	$0.602\pm 0.044$	$0.751\pm 0.035$

Table 3: Burned area mapping performance across 793 benchmark chips containing 1,162 distinct MTBS fires. Results are mean

\pm

standard deviation across three spatially blocked cross-validation folds.

Figure 6: Examples of multi-fire benchmark chips. Columns show (a) the original target-fire label, (b) the annual reference containing all same-year fires, and predictions from (c) Tessera and (d) AlphaEarth. Numbers in (b) indicate the number of same-year MTBS fires intersecting each chip. Dark grey indicates correct burn detection, blue omission error, and red commission error.

Figure 7: Fire-level burned area recall by ignition month for Random Forest models using Tessera and AlphaEarth fire-year embeddings. Boxplots show recall across MTBS fire events; numbers indicate fires per month.

Figure 8: Zero-shot wall-to-wall burned area mapping across California in 2021. (a) Combined MTBS and CAL FIRE burned area reference and (b) Tessera prediction, shown as the fraction of burned area within 0.72 km grid cells. A nonlinear color normalization is used in (a–b) to improve visibility of low burned area fractions at statewide scale. (c) Spatial distribution of Tessera prediction polygons unmatched by either MTBS or CAL FIRE, shown as counts per 15 km grid cell after excluding polygons within 300 m of a reference perimeter. Symbols show the 50 largest manually reviewed unmatched polygons: 22 were confirmed as genuine fires absent from both reference inventories and 28 as non-fire detections.

Product		Severity 1 as unburned			Severity 1 excluded
	OE $\downarrow$	CE $\downarrow$	F1 $\uparrow$	IoU $\uparrow$	CE $\downarrow$	F1 $\uparrow$	IoU $\uparrow$
Tessera	3.0%	14.6%	0.908	0.831	5.9%	0.955	0.914
MCD64A1	11.8%	15.8%	0.862	0.757	6.4%	0.908	0.832
GABAM	45.8%	7.0%	0.684	0.520	4.3%	0.692	0.529

Table 4: Statewide pixel-level performance of burned area products in California in 2021 against the combined MTBS–CAL FIRE reference. Reference data comprise 37 MTBS fires and 386 CAL FIRE fire perimeters. Metrics are reported with MTBS severity class 1 (“unburned to low”) either treated as unburned or excluded from evaluation. Omission error (OE) is unchanged between treatments and is therefore reported once.

Figure 9: Fire detection and delineation across fire sizes in California in 2021. (a) Detection rate for 386 CAL FIRE fires by size class; (b) mean fire-level IoU among detected fires; and (c) representative examples comparing the reference with Tessera, GABAM, and MCD64A1. Black indicates correct burn detection, blue omission, and orange commission;

P

and

R

denote precision and recall.

Evaluation set	$n$ fires	Micro F1 $\uparrow$	Micro IoU $\uparrow$	Macro F1 $\uparrow$	Macro IoU $\uparrow$
2024	29	0.881	0.787	$0.849\pm 0.105$	$0.750\pm 0.137$
2025	59	0.883	0.790	$0.820\pm 0.157$	$0.718\pm 0.180$
Combined	88	0.882	0.789	$0.829\pm 0.142$	$0.728\pm 0.168$

Table 5: Cross-continental burned area segmentation on 88 European EMSR fires from 2024–2025 using a Tessera model trained on US fires from 2018–2021. Micro metrics aggregate pixels across fires; macro metrics are mean

\pm

standard deviation across fires.

Figure 10: Per-fire Tessera lightweight U-Net predictions for 15 European EMSR fires from 2024–2025. Examples span the observed performance range and are ordered by decreasing IoU. Black indicates correct burn detection, blue omission, and orange commission error; bold values show per-fire IoU.

Figure 11: Qualitative error analysis for selected low-performing European EMSR fires. For each event, a Sentinel-2 false color composite (left) is paired with a Tessera Light U-Net error map (right): true positives (white), true negatives (light green), false positives (red), and false negatives (blue). Many false positives (red) coincide with visible burn scars outside the EMSR evaluation perimeter.

Figure 12: Fire timing prediction from annual embeddings. (a–b) Predicted vs observed ignition day of year for Tessera and AlphaEarth on fires with burned area recall

>0.8

. (c) Continental transfer of Tessera timing prediction to European EMSR events. (d–e) Mean absolute error and

\pm

30 day accuracy as a function of burn detection recall threshold.

Figure 13: Effect of multi-year temporal context on burned-class F1 for Tessera and AlphaEarth. Points show means across three spatial cross-validation folds; error bars indicate

\pm

1 standard deviation.

E_{-1}

,

E_{0}

, and

E_{+1}

denote the years before, containing, and after the fire, respectively, and

\Delta=E_{0}-E_{-1}

. Full details are provided in Section S9 .

Figure S1 : Examples of HLS Burn Scars benchmark labels and their refinement using MTBS burn-severity data. For each example, columns show the original benchmark burned area mask, the corresponding MTBS burn-severity mosaic, and the refined binary burned area mask used in this study. MTBS severity classes 2 (low), 3 (moderate), and 4 (high) were treated as burned, while classes 0 (unburned), 1 (unburned to low), and 5 (increased greenness) were treated as unburned. The examples illustrate how severity-based refinement removes unburned and spectrally ambiguous areas contained within the original benchmark masks.

Figure S2 : Nonlinear UMAP projections for a sample of 10 randomly selected chips. For each embedding product, UMAP was fitted jointly to the pre-fire (

E_{-1}

) and fire-year (

E_{0}

) embeddings. Points are coloured by burn status; markers show the burned and unburned centroids.

Figure S3 : PCA projections for a sample of 10 randomly selected chips. For each embedding product, PCA was fitted jointly to the pre-fire (

E_{-1}

) and fire-year (

E_{0}

) embeddings. Points are coloured by burn status; markers show the burned and unburned centroids.

Input	Model	Folds	Accuracy	F1	IoU	Precision	Recall
Tessera	LogReg	3	$0.947\pm 0.003$	$0.791\pm 0.013$	$0.655\pm 0.017$	$0.704\pm 0.016$	$0.903\pm 0.015$
	RF	3	$0.971\pm 0.002$	$0.870\pm 0.014$	$0.771\pm 0.023$	$0.857\pm 0.013$	$0.884\pm 0.019$
	Light U-Net	3	$0.980\pm 0.002$	$0.911\pm 0.011$	$0.837\pm 0.019$	$0.897\pm 0.014$	$0.927\pm 0.010$
	U-Net	3	$0.981\pm 0.001$	$0.916\pm 0.008$	$0.845\pm 0.014$	$0.903\pm 0.012$	$0.930\pm 0.014$
AlphaEarth	LogReg	3	$0.728\pm 0.021$	$0.386\pm 0.025$	$0.239\pm 0.020$	$0.258\pm 0.021$	$0.769\pm 0.017$
	RF	3	$0.892\pm 0.004$	$0.585\pm 0.013$	$0.414\pm 0.013$	$0.508\pm 0.012$	$0.691\pm 0.014$

Table S1 : Single-fire burned area delineation on 559 chips with more than 80% valid pre- and post-fire HLS coverage, evaluated with spatially blocked three-fold cross-validation. Metrics are burned-class scores from full-chip pooled evaluation (fold mean

\pm

std). MTBS severity classes 1 and 5 were counted as unburned. Tessera, AlphaEarth, and post-fire HLS results were re-evaluated on this cohort; the pre- and post-fire HLS baseline was trained on it originally.

Figure S4 : Fire-level burned area recall against ignition date: (a) Tessera and (b) AlphaEarth with the Random Forest; (c) Tessera and (d) AlphaEarth with the Light U-Net. The models were fitted on the same 1,031 MTBS fire events. Faint grey points are individual fires and the rug beneath each axis shows their timing distribution; black points are observed mean recall in 30-day bins. The dashed grey line is the unadjusted GAM of recall on ignition timing alone; the coloured line is the adjusted GAM, with a 95% ecoregion cluster bootstrap interval as the band. Adjusted curves are standardized average marginal predictions: predicted recall averaged over all fires with fire size, severity, incident type, pre-fire land cover, biome and cross-validation fold held at their observed values.

Approach	Representation	Precision $\uparrow$	Recall $\uparrow$	F1 $\uparrow$	IoU $\uparrow$
Seydi (2025)	AlphaEarth	$0.806\pm 0.211$	$0.749\pm 0.231$	$0.735\pm 0.186$	$0.612\pm 0.220$
This study	Tessera	$\mathbf{0.824\pm 0.113}$	$\mathbf{0.894\pm 0.210}$	$\mathbf{0.828\pm 0.180}$	$\mathbf{0.737\pm 0.195}$

Table S2 : Contextual comparison with Seydi (2025) on a European EMSR subset of 17 fires (

n=17

). Both methods are trained on US fires and evaluated on the same European EMSR subset; metrics are reported as mean

\pm

standard deviation across fires. Seydi results are taken from the published AlphaEarth transfer experiment.

Approach	mIoU $\uparrow$	m-F1 $\uparrow$
Annual embedding approach (single-fire subset)
Tessera + LogReg	0.778	0.866
Tessera + RF	0.850	0.915
Tessera + Light U-Net	0.922	0.959
AlphaEarth + Light U-Net	0.792	0.875
Our post-fire HLS baseline

Table S3 : Contextual comparison with published PANGAEA HLS Burn Scars benchmark results. Metrics are mean class scores. Published values are reported from the original benchmarks over the full chip set and are not strictly head-to-head with our single-fire subset evaluation; our own post-fire HLS U-Net, evaluated on both sets, quantifies the offset between them.

Figure S5 : Effect of explicit multi-year temporal context on burned area classification using Tessera and AlphaEarth embeddings. Bars show mean (a) burned-class F1 and (b) burned-class IoU across three spatial cross-validation folds; numbers above the bars give the corresponding means, and error bars indicate one standard deviation. Models were trained on balanced burned and unburned pixel samples and evaluated using 10,000 uniformly sampled valid pixels from each held-out chip.

E_{-1}

,

E_{0}

, and

E_{+1}

denote the pre-fire, fire-year, and post-fire annual embeddings, respectively, and

\Delta=E_{0}-E_{-1}

.

Features	Tessera F1	Tessera IoU	Tessera AUROC	AlphaEarth F1	AlphaEarth IoU	AlphaEarth AUROC
$E_{0}$	$0.840\pm 0.011$	$0.725\pm 0.016$	$0.973\pm 0.002$	$0.583\pm 0.041$	$0.412\pm 0.040$	$0.908\pm 0.007$
$E_{0}-E_{-1}$	$0.835\pm 0.011$	$0.716\pm 0.016$	$0.963\pm 0.002$	$0.676\pm 0.026$	$0.511\pm 0.029$	$0.930\pm 0.004$
$[E_{-1},E_{0}]$	$0.849\pm 0.012$	$0.737\pm 0.018$	$0.973\pm 0.002$	$0.613\pm 0.039$	$0.443\pm 0.040$	$0.915\pm 0.006$
$[E_{-1},E_{0},\Delta]$	$\mathbf{0.857\pm 0.009}$	$\mathbf{0.750\pm 0.014}$	$0.974\pm 0.002$	$\mathbf{0.731\pm 0.022}$	$\mathbf{0.577\pm 0.027}$	$0.949\pm 0.004$
$[E_{-1},E_{0},E_{+1}]$	$0.855\pm 0.011$	$0.747\pm 0.017$	$0.975\pm 0.002$	$0.682\pm 0.023$	$0.518\pm 0.026$	$0.944\pm 0.002$
$[E_{-1},E_{+1}]$	$0.676\pm 0.013$	$0.510\pm 0.015$	$0.940\pm 0.003$	$0.599\pm 0.025$	$0.428\pm 0.025$	$0.921\pm 0.001$

Table S4 : Effect of explicit multi-year temporal context on burned-area classification using Tessera and AlphaEarth embeddings. Random Forest models were evaluated using 10,000 uniformly sampled valid pixels per validation chip under spatially blocked three-fold cross-validation. Values are fold mean

\pm

standard deviation.

E_{0}

denotes the fire-year embedding,

E_{-1}

the preceding year,

E_{+1}

the following year, and

\Delta=E_{0}-E_{-1}

. Bold values indicate the highest F1 and IoU for each embedding product.

Severity	$n$ burned	Tessera FNR	AlphaEarth FNR	Gap (AE $-$ Tessera)
Low	16.8 M	10.8%	29.0%	$+18.2$ pp
Moderate	4.9 M	1.3%	6.3%	$+5.1$ pp
High	1.6 M	0.1%	0.8%	$+0.6$ pp

Table S5 : False-negative rates by MTBS burn-severity class for Tessera and AlphaEarth using the Lightweight U-Net. Values are aggregated across all 793 benchmark chips using out-of-fold predictions from the annual fire mapping experiment (Section 4.3 ). False-negative rates (FNR) are calculated over burned pixels within each severity class; the final column reports the AlphaEarth–Tessera difference in percentage points.

Severity	Tessera FNR	AlphaEarth FNR	Gap (AE $-$ Tessera)
Low	12.2%	28.3%	$+16.1$ pp
Moderate	3.8%	9.5%	$+5.7$ pp
High	2.1%	3.7%	$+1.6$ pp

Table S6 : False-negative rates by MTBS burn-severity class for Tessera and AlphaEarth using logistic regression. Values are aggregated across all 793 benchmark chips using out-of-fold predictions from the annual fire mapping experiment (Section 4.3 ). False-negative rates (FNR) are calculated over burned pixels within each severity class; the final column reports the AlphaEarth–Tessera difference in percentage points.

Analysis	Cohort and evaluation
Published-transfer comparison (Section S7 )	Same US-trained Tessera model evaluated without retraining on 17 EMSR fires from the published AlphaEarth comparison; 15 belong to the main cohort and two are additional 2022 fires
PANGAEA comparison (Section S8 )	Single-fire subset: 588 chips; official 394/194 split; original HLS Burn Scars masks; full-chip, mean-class scoring
Multi-year context ablation (Section S9 )	793 chips; six temporal configurations; same three-fold spatial cross-validation as annual fire mapping; RF trained on balanced samples and evaluated using 10,000 uniformly sampled valid pixels per validation chip

Table S7 : Overview of supplementary-only comparisons and ablations. Complete protocols are provided in the linked sections.

Annual Earth-observation embeddings encode wildfire disturbance and support simplified burned area mapping

Abstract

Figures & tables

Remote Sensing Image Segmentation

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

Earth embeddings are decomposable records of the land surface

Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping

Abstract

Remote Sensing Image Segmentation

Explore similar work

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

Earth embeddings are decomposable records of the land surface

Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping