Earth embeddings are decomposable records of the land surface

2.3 Feature families date disturbances on six continents

• Features selected on some events dated 85% of held-out disturbances to their expected windows; small fires and crop separation differed between Tessera and AlphaEarth.The tests covered 3,940 mapped events across six continents; small, fast-recovering fires were detectable in Tessera but not AlphaEarth, while corn and soy were better separated in AlphaEarth.Source: §1 · p. 2, Figure 3 · p. 7, Figure 6 · p. 10

2.1 Earth embeddings contain ecological features

• A sparse autoencoder decomposes each pixel’s embedding into features; on the Dixie Fire, 32 selected features together separated burned from unburned pixels with AUC 0.914.The family was selected using post-fire burned-minus-unburned activation contrast and also peaked in the first year after three other fires.Source: §2.1 · p. 3, Figure 1 · p. 4

2.2 Features carry single meanings

• A corn feature was active on pixels that were 91% corn, versus 45% across the tile, and corn and soy features tracked recorded crop rotation between years.By contrast, raw embedding dimension 47 ordered classes from corn through soy to developed land, with its extreme values dominated by different classes.Source: §2.2 · p. 3, Figure 2 · p. 5

2.4 Feature timing follows phenology and vegetation recovery

• Feature-family activation peaked in the calendar year containing the first post-disturbance growing season, reflecting event timing relative to the growing season and vegetation response.Only 2% of western US wildfires peaked in the event year, compared with 80% of southeastern US prescribed fires; most held-out tiles for two storms across a calendar boundary peaked in the predicted year.Source: §2.4 · pp. 6–7, Figure 4 · p. 8

2.5 Features record severity of an event but not its cause

• Removing the 32-feature fire family lowered burn-probe accuracy by 19 percentage points, while removing random features had no measurable effect; features distinguished fire causes only when resulting surfaces differed.The embeddings did not distinguish windthrow, bark-beetle damage and harvest when those disturbances left similar surface conditions.Source: §2.5 · p. 7, Figure 5 · p. 9
• Annual embeddings did not resolve disturbance timing within a year, and the study found no early-warning signal in pre-event feature activation.For held-out units, the pre-event effect was +0.014, with a confidence interval that included zero.Source: §2.5 · pp. 7, 9

2.6 Embedding contents differ between models

• The two products differed on identical events: Tessera detected small, fast-recovering fires, while AlphaEarth separated corn and soy more strongly and retained more disturbance signal two years later.For 240 western US megafires, family activation peaked in the first post-fire growing season for 216 in Tessera and 84 in AlphaEarth.Source: §2.6 · p. 10, Figure 6 · p. 10

3 Discussion

• The embeddings record abrupt canopy change, resolve its severity, and date it to the calendar year containing the first post-disturbance growing season.The authors say the findings are consistent with the decomposition capturing real structure rather than method artifacts.Source: §3 · p. 11
• Tessera and AlphaEarth retain different phenomena, and their comparison suggests training objectives may shape embedding content.The link between objective and retention is correlational because the comparison covers only one pair of models.Source: §3 · p. 11
• Small fires that recover quickly illustrate why benchmark performance may not indicate whether an embedding contains a user's target phenomenon.The authors recommend checking for the target phenomenon at the required temporal resolution before building a monitoring or compliance system.Source: §3 · p. 11
• Validation rates may understate how faithfully embeddings record events if reference errors are largely independent of the embeddings.That assumption may fail where reference products and embeddings derive from similar satellite data.Source: §3 · p. 11
• The analysis verifies features at individual-event and severity levels, while causal attribution, within-year timing, and severity-class boundaries remain open.The method needs only embeddings and a reference product for the phenomenon, and can apply to closed as well as open products.Source: §3 · p. 11

4.1 Embedding products

• The analysis uses frozen Tessera and AlphaEarth annual 10 m embeddings on a 0.1° tile grid, without model finetuning.Tessera embeddings are 128-dimensional and AlphaEarth embeddings are 64-dimensional.Source: §4.1 · p. 11

4.2 Sparse dictionaries

• TopK sparse autoencoders encode embeddings with a limited number of active features and decode them to reconstruct the input; the main dictionaries use matched 8× expansion.Dictionaries are trained per region without labels, reference products, or event dates, achieve median held-out R2 of 0.9, and have fewer than 3% inactive features on average.Source: §4.2 · p. 12

4.3 Feature screening

• A feature is retained when its in-window contrast exceeds a dictionary-relative threshold, earlier contrasts stay below half that threshold, and later contrast falls below half its in-window value.Changing the threshold multiplier from a quarter to four times its value changes family sizes but not interpretations.Source: §4.3 · p. 12

4.4 Validation corpus and scoring

• Features are screened on one subset of units and scored on held-out units, using event windows or specified crop-classification accuracy thresholds.Crop tiles are screened on 2018–2021 and scored on 2022–2024; a crop tile validates at accuracy of 0.70 for the crop atlas or 0.80 for crop rotation.Source: §4.4 · p. 12

4.5 Task nulls and uncertainty

• Disturbance nulls draw same-sized random feature sets, while tile-clustered bootstraps quantify uncertainty in validation rates and differences.The null uses 200 draws; confidence intervals use 10,000 tile resamples, and the crop classification analysis has no peak-offset null.Source: §4.5 · pp. 12–13

4.6 Reference data products

• Reference products supply event dates, disturbance maps, and crop labels for settings across regions, including fire, forest change, landslide, water, and crop data.The methods draw on MTBS, NBAC, Geoscience Australia, EFDA, and other public products.Source: §4.6 · p. 13

4.7 Cross-model comparison

• Cross-model comparisons measure disturbance signal over noise, post-peak retention, and corn–soy separation using distances between embedding means.AlphaEarth embeddings are rescaled to Tessera’s variance for these comparisons.Source: §4.7 · p. 13

• Embedding data and the listed reference datasets, including MTBS, NBAC, EFDA, and USDA CDL, are publicly available.Source: Data availability · p. 13

• Event-dated reference products record disturbance occurrence dates, including fire, storm, and rainfall-triggered landslide inventories.The listed sources include MTBS, NBAC, Geoscience Australia, and FORWIND.Source: Supplementary Note S1 Reference data… · p. 20
• Detection-dated products record the first annual composite showing disturbance or new water, while RADD alerts date radar detections to the day.Source: Supplementary Note S1 Reference data… · p. 20
• Crop maps provide class labels, while mining and solar datasets, WorldCover, and Copernicus DEM supply non-vegetation and probe-control inputs.The severity input is a Sentinel-2-derived composited dNBR layer.Source: Supplementary Note S1 Reference data… · p. 20
• Event and detection dates can differ by up to a year, so offsets are kept separate by date type and expected windows are fixed by setting.Rasters are reprojected onto the embedding tile grid at acquisition to avoid post-masking resampling.Source: Supplementary Note S1 Reference data… · p. 20
• For Fig. 1, burned area is defined by ΔNBR > 0.27; the Sentinel-2 composite covers 69.4% of the Dixie tile, and AUC is computed over covered pixels.Source: Supplementary Note S1 Reference data… · p. 20

Embeddings

• Tessera embeddings are 128-dimensional and AlphaEarth embeddings are 64-dimensional, both at 10 m, and the analysis uses frozen embeddings without finetuning.AlphaEarth embeddings are reprojected to the same 0.1° tile grid as Tessera.Source: Supplementary Note S2 Embeddings and… · p. 20

Dictionary

• The reference configurations use TopK sparse autoencoders with an eightfold expansion for both embedding products.Dictionaries are trained per region; the three Europe atlas settings share a pan-European dictionary.Source: Supplementary Note S2 Embeddings and… · p. 21
• The TopK autoencoder reconstructs embeddings from a sparse set of learned features, keeping only the k largest preactivations.Sparsity is imposed by TopK rather than an L1 penalty.Source: Supplementary Note S2 Embeddings and… · p. 21

Feature purity at scale

• On 100 Corn Belt tiles, fixed corn and soy features reach median F1 0.91 and 0.92; fixed raw-dimension rules reach 0.90 and 0.87 (Fig. S1).The features outperform the fixed raw-dimension rule on 71 tiles for corn and 98 for soy.Source: Supplementary Note S2 Embeddings and… · p. 21, Figure S1 · p. 21

Measurement

• The contrast compares a feature’s mean activation on disturbed pixels with its mean activation on undisturbed pixels in the same tile.Contrasts are indexed by offset from the event year and combined across units by their median.Source: Supplementary Note S3 Measurement and… · p. 22

Screening

• Features join a family when their contrast exceeds the anchored threshold in the expected window and remains below half that threshold before the event.Missing expected-window measurements score zero, while an absent pre-event offset imposes no constraint.Source: Supplementary Note S3 Measurement and… · p. 22

Multiplier sensitivity

• Changing the screening multiplier leaves validation rates within three points of reported values in most disturbance settings.Family sizes respond strongly to the threshold, but the held-out outcomes generally do not.Source: Supplementary Note S3 Measurement and… · pp. 22–23, Table S1 · p. 23

Seed sensitivity

• Across five retrained Dixie dictionaries, western US validation rates range from 0.84 to 0.92, far above random-family medians of 0.12–0.15.Individual features vary between seeds, while family mean decoder directions remain similar, with cosine 0.93–0.96.Source: Supplementary Note S3 Measurement and… · p. 23

Full results

• Table S2 reports reference products, date types, expected windows, held-out unit counts, validation rates, and null baselines across 16 settings.Disturbance validation is the share of held-out units peaking in the expected window; crop validation is held-out classification accuracy.Source: Supplementary Note S4 Event corpus… · p. 23
• Crop settings have no peak-offset null because their units are scored by held-out classification accuracy.A peak-offset null is undefined for that crop-scoring procedure.Source: Supplementary Note S4 Event corpus… · p. 23

Random-family nulls

• Restricting random-family draws to live features outside the screened family lowers the null in all but one setting, where it changes by 0.01.The reported Table S2 nulls are therefore conservative.Source: Supplementary Note S4 Event corpus… · p. 23

Western US fire

• The western US fire test includes 207 of 242 qualifying large fires, scoring a 32-feature family screened on the separate Dixie Fire.Fires require at least 20,000 acres and a perimeter covering at least 15% of a tile.Source: Supplementary Note S5 Unit selection… · pp. 23–24

Fire settings

• Fire tests also hold out later southeastern US and boreal fires, and most southeastern Australia Black Summer fires, after screening on earlier or separate events.Boreal testing uses AlphaEarth because Tessera did not fully cover Canada for the tested period.Source: Supplementary Note S5 Unit selection… · pp. 24–25

Europe atlas and windthrow

• The Europe atlas tests harvest, wind or bark-beetle damage, and fire across 436 tiles in 36 countries; windthrow families are screened on Friederike tiles and tested on remaining or other-storm tiles.The Europe settings use one Europe-wide dictionary.Source: Supplementary Note S5 Unit selection… · p. 25

Other disturbance settings

• Other disturbance settings select held-out units from Amazon clearing, Africa and East Asia forest loss, rainfall landslides, and newly formed permanent water.The RADD clearing test dates plots by median alert date and tests all 17,014 plots after the family’s 2018 screen.Source: Supplementary Note S5 Unit selection… · pp. 25–26

Crop atlas

• The crop atlas assigns held-out pixel-years among seven competing classes and reports pooled accuracy of 0.972 for corn, 0.926 for soy, 0.795 for winter wheat, 0.841 for rice, and 0.620 for cotton.The unweighted mean accuracy is 0.831.Source: Supplementary Note S5 Unit selection… · p. 26

Crop rotation

• In the Corn Belt rotation test, 95 of 100 tiles reach at least 0.80 held-out corn-versus-soy accuracy using features alone.Features are screened on 2018–2021 crop labels and evaluated on 2022–2024.Source: Supplementary Note S5 Unit selection… · p. 26

Supplementary Note S6 Families for non-vegetation surfaces

• The analysis extends the dated disturbance tests to surfaces that are not vegetation.The authors frame this as testing whether embeddings contain features for non-vegetated surfaces.Source: Supplementary Note S6 Families for… · p. 26
• Mining-land pixels are drawn from five classes in the global mining land use dataset, with nearby labeled polygons excluded from the surroundings.The study samples up to 30 large sites per world region and reads Tessera embeddings within two years of the label year.Source: Supplementary Note S6 Families for… · p. 26
• Solar-farm target pixels come from eroded outlines of operating US ground-mounted farms, while surroundings lie 100 m to 2 km from mapped farms.The 300 sampled tiles have at least 25 ha of farm area; embeddings are from Tessera 2024.Source: Supplementary Note S6 Families for… · p. 26
• A feature family is screened on a random fifth of sites using target-versus-surroundings activation contrast, then tested on held-out sites with an unchanged cutoff.A site validates at balanced accuracy 0.70; the random-family null uses 200 families of the same size, and regional transfer uses separate screening and scoring regions.Source: Supplementary Note S6 Families for… · pp. 26–27
• Both mining-land and solar-farm feature families validate on most held-out sites, far above random families, and transfer across regions.The reported accuracy is median balanced accuracy; AUC is median per-site AUC of family activation for target versus surroundings.Source: Supplementary Note S6 Families for… · p. 27, Table S4 · p. 27

Supplementary Note S7 Feature counts are instrument-relative

• A feature counts as verified only when it peaks in the year after a fire, and the anchored and fixed bars coincide at the originally calibrated dictionary configuration.The multiplier is held unchanged as dictionary capacity varies.Source: Supplementary Note S7 Feature counts… · p. 27

Supplementary Note S8 Cross-model comparability

• Tessera and AlphaEarth comparisons match embedding scale, dictionary expansion relative to embedding dimension, and event, mask, screen, and expected-window choices.The cross-model dissociation uses raw embeddings.Source: Supplementary Note S8 Cross-model… · p. 27

Supplementary Note S7 Feature counts are instrument-relative

• The anchored screening bar yields growth in verified burn-history features in all four measured settings, while the fixed bar is at or below 1 in three settings and reaches 1.6 at k = 32.The fixed bar tightens from 13.8× to 660× each dictionary’s median in-window contrast as capacity expands.Source: Supplementary Note S7 Feature counts… · p. 27, Figure S2 · p. 28

Supplementary Note S8 Cross-model comparability

• The raw-embedding signal measures use the distance between mean disturbed and undisturbed pixel embeddings by year; persistence is computed only when peak signal exceeds one tenth of the noise floor.The pre-event noise floor is 1.0.Source: Supplementary Note S8 Cross-model… · p. 28

Supplementary Note S9 Specificity and invariance

• The unchanged western US fire family responds to European fire but not harvest or windthrow, with 38% of fire units exceeding the random-set 95th percentile versus 3% of harvest and none of 89 windthrow units.Its European fire response is weaker than on the Dixie fire; the table reports the unchanged 32-feature family scored through the Dixie dictionary.Source: Supplementary Note S9 Specificity and… · p. 28, Table S5 · p. 29
• The transfer tests apply families screened in one region to unseen regions for rice, winter wheat, and fire.The reported transfers are Arkansas rice to California’s Sacramento Valley, US southern Plains wheat to several European countries, and western US fire to European burns.Source: Supplementary Note S9 Specificity and… · p. 29
• Rice and winter-wheat families transfer to new regions without loss in AUC: rice 0.95 to 0.96 and wheat 0.80 to 0.80.The concepts nearly coincide along their family axes despite being 2.7 and 3.8 standard deviations apart along the most separating raw-embedding direction.Source: Supplementary Note S9 Specificity and… · p. 29, Table S6 · p. 29

Supplementary Note S10 Feature ablation

• The probes predict burn history, terrain aspect, and tree-versus-grass cover using held-out pixels split by spatial checkerboard blocks.The latter two targets test specificity because they are co-resident in the embedding and are not functions of fire history.Source: Supplementary Note S10 Feature ablation · pp. 29, 31
• The figure guide describes cross-region activation patterns for rice, winter wheat, and fire, including raw embedding projections and summed family activation.It shows the concepts and their surroundings in each region, with groups also plotted along a raw-embedding direction and in activation space.Source: Figure S3 · p. 30
• The ablation subtracts a feature family’s decoder-direction contributions from frozen embeddings, then compares probe performance with 200 same-size random feature-set ablations.Both family and random-set ablations are applied to burn-history, terrain-aspect, and tree-versus-grass probes.Source: Supplementary Note S10 Feature ablation · p. 31
• The precursor screen reverses the disturbance screen’s pre-event-quiet condition and counts only signals at Y −2 or earlier as precursor evidence.Because the reference atlas dates disturbance at the first clear post-event composite, Y −1 may already represent the event; controls are matched within tiles by distance to prior disturbance, with harvest as a specificity control.Source: Supplementary Note S10 Feature ablation · p. 31