cs.LGOct 23, 2026
SaveEarth embeddings are decomposable records of the land surface
Organizations: Department of Computer Science and Technology, University of Cambridge, William Gates Building, Cambridge CB3 0FD, UK
Abstract
Ecological monitoring, carbon accounting and biodiversity assessment increasingly rely on annual Earth embeddings--learned vectors that compress a year of satellite observation for every ten-meter pixel on Earth, consumed in place of the imagery itself. What these vectors mean has not been established. Here we decompose two widely used embedding products and show that they contain individual ecological features, including fire, forest clearing and crop type, validated against independent reference data on events across six continents. The recovered concepts date abrupt canopy loss to the first post-disturbance growing season and resolve severity well enough to separate prescribed fire from wildfire. Small fires that recover quickly are detectable in one product but not the other, which instead separates crops more strongly.
Figures & tables
Fig. 1: Sparse decomposition recovers a fire from Earth embeddings. a , The Dixie Fire (California) in Sentinel-2 shortwave infrared imagery after containment. b , A year of Sentinel-1 and Sentinel-2 observations of the same area, compressed into Tessera’s 128-dimensional annual embedding (first three principal components. Inset individual 10m pixels). c , A sparse autoencoder decomposes each pixel’s embedding. Below we show a few features’ activation maps. The burn features family (32 features) is the set selected from burned-minus-unburned contrast after the fire. Other illustrative features are south-facing slopes (91% of its active pixels face south, from the Copernicus DEM, against 55% of the tile) and a meadow feature whose activation rises from 2% of forest pixels more than 300m from a meadow to 51% of pixels more than 100m inside meadows (ESA WorldCover). d , Burn severity from the dNBR spectral index. e , Mean activation of the burn family, which separates burned from unburned pixels with AUC 0.914. 96.0% of pixels with above-median activation are burned (dNBR , pixels mapped as moderately or severely burned). f , Burned-minus-unburned family activation by year since fire, for the Dixie Fire (coral) and for three Sierra Nevada fires on which the family was not screened (gray lines are North Complex 2020, Caldor 2021, Mosquito 2022).
Fig. 2: A decomposed feature identifies one land cover class. a , USDA Cropland Data Layer (CDL) classes on one Iowa tile in 2020. b , Pixels in the top and bottom 2% of raw embedding dimension 47, colored by CDL class. The bottom tail is mostly corn (93%) and the top tail mostly developed land (68%). c , Pixels where decomposed feature 904 is active, colored by CDL class. 91% are corn, against 45% across the tile overall. d , Class composition of the tile’s pixels, ranked by their value on dimension 47 and split into 20 bins of equal pixel count (left), and of pixels where feature 904 (corn) is off or active (right). The dimension orders classes from corn through soy to developed land and identifies a single class only near its ends. The feature is active on 40% of the tile, and 91% of those pixels are corn. e , Share of pixels where the corn and soy features are active in 2020 and 2021, on pixels the CDL records as corn then soy (left, 66,881 pixels) and as corn in both years (right, 33,152 pixels).
Fig. 3: Validating decomposed feature families across 3,940 units in 16 settings on six continents. a , Every unit, colored by its outcome of peaking in its setting’s expected window, early or late. Filled markers are event-dated references and open markers detection-dated, which record different physical dates. Purple squares are crop units, which are classified. b , Held-out validation rate per dated setting (filled) with its 95% confidence interval, against the median of a random-family null computed for that setting (open), with the number of held-out units ( ). Rates and nulls are computed on the same held-out units. All 14 dated settings are above their nulls at .
Fig. 4: Disturbance features peak in the first growing season that shows the disturbed surface. a , Share of units peaking in the event year, by ignition month, one curve per setting. Wilson 95% intervals are shown, point area scales with . Note each setting’s growing season (western US April–October, SE US March–October). Western US wildfire peaks in the following year at nearly every ignition month, while SE US wildfire and prescribed fire mostly peak in the event year. Months with fewer than 4 events are dropped. b , Three windstorms in the same forest type with Xavier and Friederike ten weeks apart across a calendar boundary, and Vaia after the 2018 growing season, with markers at event dates, green bands marking each storm’s first post-storm growing season, and stars the observed peak years. Friederike (23 of 23 tiles) and Xavier (6 of 6) peak in the predicted year, and Vaia in 41 of 60. c , Detection-dated settings, where the reference dates the first annual composite showing the disturbance.
Fig. 5: Removing the burn family removes what a probe uses to detect fire, and features record severity of an event but not its cause. a , Change in linear probe accuracy when features are removed from the reconstruction. Removing the 32-feature burn family costs the burn probe 19 points, while a single strong feature moves the accuracy only a little, and a matched random set of the same size (mean over 200 draws, band 95% CI) has nearly no effect. Neither ablation moves other probes (terrain aspect, tree vs grass) by more than random chance. b , All size- and region-matched SE US fires (191 wildfire, 145 prescribed burn) ranked by the burned minus unburned activation of the SE wildfire family and split into deciles. Dashed line is pooled wildfire share (57%) as baseline if activation carried no differentiating information. The four strongest tenths are 88–94% wildfire. c , 309 matched pairs of wind and bark-beetle and harvest tile-years, ranked by wind-only minus harvest-only feature activation. Every tenth sits near the pooled 50%, so the specific features do not resolve cause where the surface signals are similar.
Fig. 6: Tessera and AlphaEarth retain different phenomena on identical events. Panels use raw embeddings on identical events and masks. a , Top: small fast-recovery fires (SE US wildfire), raw signal over pre-event noise. Tessera sits at 1.56 and AlphaEarth at 1.09, at the noise floor (dotted). The identical feature screen yields 15 features in Tessera and none in AlphaEarth over the same 21 events. Bottom: crop identity (Corn Belt corn vs soy, separation of class means in units of within-class spread), Tessera 1.43 and AlphaEarth 3.19. b , Fraction of disturbance signal remaining two years after identical events, per setting. AlphaEarth retains more in all three (ratios 2.8, 3.5, 1.7). c , Each model’s family-peak offset from the fire year over 240 western US megafires both screened. Green band marks the expected window.
Fig. S1: One fixed feature against one fixed raw-dimension rule. For each of 100 Corn Belt tiles, the F1 with which the corn and soy features of Fig. 2 , read as active or inactive, identify their crop (vertical), against the best single raw embedding dimension with one end and threshold chosen on the tile of Fig. 2 (horizontal). Points above the line are tiles where the feature does better.
| Setting | 7.25 | 14.5 | 29 | 58 | 116 | |
|---|---|---|---|---|---|---|
| Western US fire | 207 | 0.884 (74) | 0.894 (62) | 0.889 (32) | 0.894 (11) | 0.841 (4) |
| SE US wildfire | 154 | 0.701 (14) | 0.773 (15) | 0.792 (15) | 0.896 (15) | 0.890 (14) |
| SE US prescribed fire | 109 | 0.761 (8) | 0.752 (11) | 0.798 (12) | 0.817 (10) | 0.798 (11) |
| Boreal fire | 309 | 0.961 (20) | 0.955 (19) | 0.939 (20) | 0.955 (17) | 0.948 (15) |
| SE Australia fire | 274 | 0.661 (37) | 0.664 (33) | 0.624 (30) | 0.639 (26) | 0.631 (20) |
| Europe harvest | 826 | 0.821 (67) | 0.818 (67) | 0.832 (66) | 0.809 (54) | 0.825 (42) |
Table S1: Sensitivity of the held-out validation rate to the screening multiplier. For each disturbance setting the family is rederived from its screening tile-years with the multiplier in Supplementary Equation ( S3 ) set to the column’s value, and the same held-out units are re-scored under the unchanged expected window. Cells give the validation rate with the family size in parentheses
| Setting | Product | Embeddings | Dates | Window | Validated | Null (median [95% int.]) | |
|---|---|---|---|---|---|---|---|
| Western US fire | MTBS | Tessera | event-dated | {Y 1} | 207 | 0.889 | 0.140 [0.024, 0.406] |
| SE US wildfire | MTBS | Tessera | event-dated | {Y, Y 1} | 154 | 0.792 | 0.312 [0.155, 0.656] |
| SE US prescribed fire | MTBS | Tessera | event-dated | {Y, Y 1} | 109 | 0.798 | 0.248 [0.101, 0.505] |
| Boreal fire | NBAC | AlphaEarth | event-dated | {Y 1} | 309 | 0.939 | 0.196 [0.042, 0.523] |
| SE Australia fire | GA Bushfire | AlphaEarth | event-dated | {Y, Y 1} | 274 | 0.624 | 0.369 [0.201, 0.639] |
| Europe harvest | EFDA | Tessera | detection-dated | {Y, Y 1} | 826 | 0.832 | 0.301 [0.113, 0.529] |
Table S2: Full results over 16 settings, 3,940 reference units. Validation is the share of held-out units whose family peak falls in the expected window (disturbance settings) or held-out classification accuracy (crop settings). Nulls are same-size random-family draws (median and 95% interval over 200 draws)
| Setting | Null (as implemented) | Null (live, non-family) | Observed |
|---|---|---|---|
| Western US fire | 0.14 | 0.09 | 0.89 |
| SE US wildfire | 0.31 | 0.30 | 0.79 |
| SE US prescribed fire | 0.25 | 0.26 | 0.80 |
| Boreal fire | 0.20 | 0.17 | 0.94 |
| SE Australia fire | 0.37 | 0.32 | 0.62 |
| Europe harvest | 0.30 | 0.19 | 0.83 |
Table S3: Random family null is conservative. Null validation rate (median over 200 draws) as implemented, with random feature sets drawn from the whole dictionary including family features, and with draws restricted to live features outside the family, against the observed rate. A feature is live if its contrast is nonzero in at least one held-out unit; dead features are 0.0–3.5% of each dictionary. Restricting draws lowers the null in 12 of 14 settings.
| Family | Screened / held out | Pass | Null | Accuracy | AUC | Cross-region | |
|---|---|---|---|---|---|---|---|
| Mining land | 23 | 43 / 174 | 0.82 | 0.07 (0.31) | 0.81 | 0.94 | 0.82 ( ) |
| Solar farms | 14 | 60 / 239 | 0.75 | 0.03 (0.31) | 0.87 | 0.97 | 0.82 ( ) |
Table S4: Families for non-vegetation surfaces. Accuracy: median balanced accuracy. AUC: median per-site AUC of the family’s activation, target against surroundings.
Fig. S2: Verified feature counts under two screening bars. a , Externally verified burn-history features against dictionary expansion ( to the 128-dimensional embedding), for three sparsity levels , under a bar anchored to each dictionary’s contrast distribution (solid) and a fixed absolute bar (dashed), on the same dictionaries with the same verification. b , The fixed bar, in multiples of each dictionary’s median in-window contrast, tightens from to (Tessera in gray, AlphaEarth in blue). c , Growth ratio (verified features at largest over smallest capacity) for the four measured settings, with the dotted rule marking no growth. The anchored bar shows growth in 4 of 4 settings, and the fixed bar sits at or below 1 in three of the four and reaches 1.6 at .
| European fire | Harvest | Windthrow | |
|---|---|---|---|
| Units | 184 | 69 | 89 |
| Peak contrast, median (max) | 0.059 (0.51) | 0.009 (0.05) | 0.010 (0.06) |
| Random-set peak, median | 0.021 | 0.010 | 0.017 |
| Percentile vs random, median | 87th | 45th | 42nd |
| Above random 95th percentile | 38% | 3% | 0% |
| Peaks in window | 68% | 36% | 12% |
Table S5: Western US burn family on European disturbances. The unchanged 32-feature family, scored through the Dixie dictionary. Each unit is a tile-year contrasted against the tile’s undisturbed forest; percentiles are relative to 200 random 32-feature sets from the same dictionary, per unit. Window is the share peaking in . Windthrow is the corpus windthrow set (held-out Friederike, Xavier and Vaia tiles).
Fig. S3: A feature family found in one region responds to the same concept in another. Each row takes one feature family in a region, and applies in a new region: rice (Arkansas to Sacramento Valley, California), winter wheat (US southern Plains to France, Germany, Poland and Hungary) and fire (California to Europe). Each row has four groups with equal samples, namely the concept in each region (e.g. US wheat, European wheat) and its surroundings in each region (other cropland for the crops, unburned forest for fire). a, d, g, Raw embeddings of the four groups on their first two principal components. b, e, h, The same points colored by the family’s summed activation (log scale, gray is inactive). c, f, i, The same four groups along a single line, once in raw embedding space (top) and once in activation space (bottom). Top: position on the first principal component of the group centroids (the direction in which the groups of a differ most). Bottom: the log of the family’s summed activation. The inset repeats the family’s active curves, magnified vertically, over the same x range. Triangles mark the two concept centroids.
| Learned in tested in | AUC, learned | AUC, tested | Location | |
|---|---|---|---|---|
| Rice | Arkansas Sacramento Valley | 0.95 | 0.96 | 0.26 |
| Winter wheat | US Plains Europe | 0.80 | 0.80 | 0.39 |
| Fire | California Europe | 0.85 | 0.69 | 0.58 |
Table S6: Transfer of three feature families to a new region. AUC: family activation, concept against its surroundings (other cropland, intact forest). Location is share of raw embedding variance among pixels where the family is active that is explained by location.
More in this topic
Geospatial Representation Learning
Geospatial representation learning builds reusable embeddings of places, imagery, maps, mobility, and Earth data for spatial prediction and reasoning.
See all 133 papers