AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data

Introduction

• Global mapping remains constrained by scarce ground measurements, sparse high-quality annotations, and the overwhelming volume of geospatial data.Existing mapping efforts often trade measurement precision and spatial coverage, while highly detailed observations remain localized in space and time.Source: Introduction · p. 1, From sparse labels to maps · pp. 1, 3

From sparse labels to maps

• AEF reconciles temporally, geographically, and spatially sparse measurements from multiple raster and scalar sources into continuous embedding fields despite uneven data availability.The model separates the input support period from the conditioning valid period and produces annualized planet-scale layers over approximately 10-meter grids.Source: From sparse labels to maps · p. 3, AlphaEarth Foundations · p. 4
• Designed EO features can efficiently extrapolate labels but are often noisy, sensor-dependent, and region- or application-specific; existing learned approaches do not consistently win in scarce-data regimes.Prior learned approaches also lack some combination of multi-source modeling, temporal integration, and operationally useful spatial precision.Source: From sparse labels to maps · p. 3

AlphaEarth Foundations

• AEF is the only learned EO featurization approach reported to outperform representative designed and learned methods across broad sparse-data domains while maintaining 10-meter resolution and 64-byte representations.The 64-byte representation requires 16x less information than the next-most compact learned method.Source: AlphaEarth Foundations · p. 3
• AEF supports continuous-time summaries that can interpolate or extrapolate beyond the timestamps of its input observations without fine-tuning.Each embedding is 64 bytes and encodes temporal trajectories over a specified valid period, with embeddings constrained toward uniformity on S63.Source: AlphaEarth Foundations · p. 4

Evaluation in realistic data-scarce scenarios

• The evaluation suite contains 15 tests from 11 open datasets spanning classification, regression, and change detection, with one-, ten-, and max-trial transfer using k-nearest neighbors or linear layers.The max-trial uses hundreds rather than thousands or millions of samples, while one- and ten-shot trials probe extreme data sparsity.Source: Evaluation in realistic data-scarce… · pp. 4, 6
• ∼23.9%: AEF reduced average error magnitudes versus the next-best approach in max-trial evaluations, with reductions of ∼10.4% for ten-shot and ∼4.18% for one-shot trials.AEF consistently outperformed both designed and learned featurization methods across all trial settings, although the next-best approach varied by dataset and method.Source: Figure 1 · p. 2, Evaluation in realistic data-scarce… · p. 6

Thematic mapping

• AEF achieved the greatest error reductions on annual-period thematic mapping evaluations, including LCMAP land cover, Descals oil palm, Africa crop mask, and LCMAP land use.The evaluation suite covered 11 classification tasks spanning land use, land cover, crop detection, crop type, and species distribution, with varied semantic complexity and summary periods.Source: Thematic mapping · p. 6
• AEF demonstrated improved spatial coherence without loss of spatial precision in qualitative comparisons at 10m2 resolution.Comparisons used AEF against the next-best model or dataset and cloud-free Sentinel-2 L1C RGB imagery from the same year, though not necessarily the same timing.Source: Thematic mapping · p. 9
• All thematic mapping evaluations except Ethiopia crops had > 1.0x error reductions within the ∼90% confidence interval, indicating generality across diverse applications.AEF’s consistent performance extended beyond higher-dimensional learned embeddings that had previously been used for these tasks.Source: Thematic mapping · p. 9

Estimating biophysical variables

• AEF achieved the highest emissivity R2 at 0.72 ± 0.00, ahead of MOSAIKS at 0.69 ± 0.00, while all baselines except xy, xyz, and CCDC exceeded R2 > 0.5 for some transfer method.Emissivity is a unitless measure of surface radiation.Source: Estimating biophysical variables · p. 9
• For evapotranspiration, AEF was the only tested method with R2 > 0.2, achieving R2 = 0.58 ± 0.01 and significantly outperforming the other baselines.Evapotranspiration characterizes water loss from Earth’s land surface to the atmosphere.Source: Figure 3 · p. 8, Estimating biophysical variables · p. 9
• Before AEF, the only evapotranspiration baselines with explanatory power were composites and MOSAIKS, both simple transformations of raw satellite data.This result indicates a gap in the applicability of prior learned and designed featurization approaches.Source: Estimating biophysical variables · p. 9

Evaluation design

• The evaluations compare direct binary change classification with unsupervised deviation thresholding on LCMAP-derived land-cover and land-use change labels, excluding time-invariant controls and SatCLIP.Direct classification treats change between two summary periods as a binary label, whereas unsupervised detection thresholds continuous deviation from an expected value.Source: Change detection · p. 9

Direct classification

• 78.4% ± 1.11 BA (linear) and 79.3% ± 1.67 BA (kNN, k=3) were achieved by AEF, exceeding next-best baselines of 72.0% ± 1.28 BA and 71.5% ± 2.33 BA in land-cover and land-use change evaluations, respectively.The next-best land-cover baseline was MOSAIKS, kNN, k=3; the next-best land-use baseline was composite, kNN, k=3.Source: Change detection · p. 9

Unsupervised thresholding

• 71.3% ± 1.14 BA made AEF the top performer for unsupervised land-cover change detection, but 71.4% ± 2.08 BA trailed ViT’s 72.9% ± 1.97 BA for land-use change.For land-cover change, the comparison baseline was ViT at 67.0% ± 1.28 BA; these results are shown in Figure 3.Source: Change detection · p. 9

• Training uses over 3 billion observations from nine gridded sources and one unstructured text source, covering approximately 1.1% of Earth’s land surface.Source: Scaling source data quantity and type · p. 9
• AEF generally outperforms other learned featurizations trained on the same or fewer observations and always outperforms them with the full training set.For US trees, AEF requires approximately 100x more observations than SatCLIP, possibly because AEF receives no coordinate information and must learn climate gradients from examples.Source: Figure 4 · p. 10, Scaling source data quantity and type · p. 11
• Adding training observations improves AEF embeddings, with performance saturating between 100 million and 1 billion observations for LUCAS land use and Africa crop mask but not obviously for US trees.Source: Figure 4 · p. 10, Scaling source data quantity and type · p. 11
• AEF performs best with the full set of Optical, Radar, LiDAR, Environmental, and Annotated source groups, although additional groups yield diminishing returns.All performance differences from adding each source group are significant at α=5%, with saturation apparent after the LiDAR or Environmental groups for some evaluations.Source: Figure 4 · p. 10, Scaling source data quantity and type · p. 11

Dataset release

• Annual AEF embedding summaries are hosted as an image dataset on Google Earth Engine for use by Earth-observation practitioners.The dataset is intended to support mapping workflows that otherwise require large training datasets, compute-intensive models, and custom inference systems.Source: Global embeddings dataset · p. 11

Quantization

• 8-bit quantization reduces embedding storage by 4x with negligible performance impact.The original 32-bit floating-point embeddings are quantized to 8 bits to minimize compute and storage overhead.Source: Global embeddings dataset · p. 11

Conclusions

• AEF combines diverse geospatial records into a time-continuous embedding space by modeling temporal dynamics and cross-source relationships.Separating measurement-specific information from mutual information across sources enables compact surface representations robust to noisy and sparse Earth-imaging data.Source: Conclusions · p. 11
• AEF consistently outperforms designed and learned featurization approaches in relatively sparse data regimes across diverse biodiversity, ecology, and agriculture applications.The embeddings support efficient mapping of spatial and temporal change when large annotation corpora are unavailable.Source: Conclusions · p. 11

Data and Materials Availability

• Annualized embedding field layers from 2017–2024, evaluation datasets, and training-site locations are released under an open license for exploration and applied use.The release is intended to support further exploration and applied use.Source: Data and Materials Availability · p. 12
• Training used publicly available data from Copernicus, USGS, NASA, JAXA, C3S, and ECMWF.These sources supplied the publicly available training data used by AlphaEarth Foundations.Source: Data and Materials Availability · p. 12
• Evaluation datasets were derived from public sources including LCMAP CONUS, LUCAS Harmonized, GLanCE, cropland maps, Canadian AAFC, Ethiopian Crop Type, iNaturalist, oil-palm maps, OpenET, and AG100.The listed datasets carry public-domain, Creative Commons, Open Government Licence Canada, non-commercial, or unrestricted reuse terms as specified in the source materials.Source: Data and Materials Availability · p. 12
• Training-site selection was informed by RESOLVE Ecoregions, the Allen Coral Atlas, and the Murray Global Intertidal Change Classification.Each listed source is governed by the Creative Commons Attribution 4.0 International License.Source: Data and Materials Availability · p. 12

S1. Data sources and preprocessing; S1.10. Text sources

• AEF integrates heterogeneous Earth-observation and auxiliary data sources spanning optical, SAR, climate, LiDAR, gravity, topography, land cover, and geocoded text.All raster sources were sampled from the Earth Engine Data Catalog, while English Wikipedia served as an auxiliary text source; Table S1 describes diversity in measurement spaces, resolutions, and temporal refresh rates.Source: S1. Data sources and preprocessing · p. 19, Table S1 · p. 20, S1.10. Text sources · p. 23

S1. Data sources and preprocessing

• All raster inputs were reprojected to UTM, resampled to 10 m, standardized using per-band pretraining statistics with 6-standard-deviation clipping, and trained with valid masks rather than input masking.Millisecond acquisition timestamps were retained as reconstruction metadata unless otherwise stated.Source: S1. Data sources and preprocessing · p. 19

S1.1. Sentinel-2 (optical); S1.2. Landsat 8 & 9 (optical, thermal)

• Sentinel-2 and Landsat 8/9 provide complementary optical records, with Sentinel-2 using five selected bands and Landsat adding panchromatic and thermal information.Sentinel-2 uses harmonized TOA imagery, while Landsat uses Collection 2 Tier 1 TOA imagery and filters nighttime observations; both missions apply source-specific transformations and training-only quality masks.Source: S1.1. Sentinel-2 (optical) · p. 19, S1.2. Landsat 8 & 9 (optical, thermal) · pp. 19, 21

S1.3. Sentinel-1 (C-band SAR)

• Sentinel-1 C-band SAR preprocessing uses calibrated GRD imagery from ascending and descending orbits, log-scaled polarization and angle bands, value-range masking, metadata retention, and simulated gap artifacts for reconstruction.The SAR mission supplies cloud-penetrating observations, including in persistently cloudy regions.Source: S1.3. Sentinel-1 (C-band SAR) · p. 21

S1.4. PALSAR-2 (L-band SAR)

• PALSAR-2 adds L-band SAR observations, including HH, HV, and incidence-angle information, with radiometric terrain correction, decibel conversion, deduplication, and pass-direction metadata.Its longer wavelength is described as more penetrating through dense vegetation than Sentinel-1’s C-band.Source: S1.4. PALSAR-2 (L-band SAR) · pp. 21–22

S1.5. ERA5-Land (climate); S1.6. GEDI (LiDAR); S1.7. GRACE (gravity fields)

• Auxiliary physical measurements include ERA5-Land monthly precipitation, temperature, dewpoint, and pressure statistics; GEDI canopy-height relative-height metrics; and GRACE terrestrial-water-storage anomalies.GEDI reconstruction has mean absolute error ∼3.85m during training, while GRACE’s coarse 0.5° resolution is smoothed through additional bilinear upsampling and downsampling.Source: S1.5. ERA5-Land (climate) · p. 22, S1.6. GEDI (LiDAR) · p. 22, S1.7. GRACE (gravity fields) · pp. 22–23

S1.8. GLO-30 (topography); S1.9. NLCD (land cover)

• Topographic and land-cover supervision uses GLO-30 elevation with UTM-derived slope and aspect, plus NLCD land-cover classes matched to nearby release years where available.NLCD is included as weak supervision; it is unavailable globally and was reported to have generally positive evaluation effects without significant negative impact on other-source reconstruction quality.Source: S1.8. GLO-30 (topography) · p. 23, S1.9. NLCD (land cover) · p. 23

S1.11. Wikipedia; S1.12. GBIF

• Geocoded text combines filtered coordinate-linked Wikipedia articles with quality-controlled GBIF plant, animal, and fungi observations from 2017–2023, including taxonomy identifiers and spatial uncertainty for matched sampling.Wikipedia excludes extraterrestrial coordinates, very short articles, and list-dominated content; GBIF records require permissive licensing, bounded spatial uncertainty, and additional validity filters.Source: S1.11. Wikipedia · p. 23, S1.12. GBIF · pp. 23–24

S2.1.1. Training site selection

• Training sites combined gridded geocoded text, stratified RESOLVE ecoregions, coral reefs, and intertidal ecosystems, yielding 5,145,244 unique locations after filtering.The sampling strategy addressed source-location bias, expanded ecological coverage, and removed open-water and redundant rock-and-ice samples.Source: S2.1.1. Training site selection · pp. 24–25

S2.1.2. Adding time coordinates

• Two non-overlapping temporal support periods per site produced 10,203,798 unique (x, y, t_start, t_end) rows for source-data collection.Periods were selected to represent temporal variability while allocating geocoded observations to only one period when present.Source: S2.1.2. Adding time coordinates · p. 25

S2.1. Training Dataset

• AEF’s final pretraining dataset comprised 8,412,511 coordinates from 5,145,244 sites, two temporal periods per site, and 3,047,520,515 image frames.The dataset was designed to represent terrestrial and near-shore ecosystems across spatial, temporal, and source-availability contexts.Source: S2.1. Training Dataset · p. 24, S2.1.3. Training sample · p. 25, S2.1.2. Adding time coordinates · p. 25

S2.2.1. Simulating a continuous observation record

• The model summarizes observations with time-axial attention, sinusoidal timecodes, learned upsampling, and a spatial von Mises–Fisher bottleneck that preserves precision and parameterizes embedding-manifold smoothness.Conditional decoders reconstruct source frames at arbitrary timestamps using source metadata, while source-specific losses accommodate categorical data, misregistration, missingness, and resolution mismatch.Source: S2.2.1. Simulating a continuous… · p. 26

S2.2.2. Learning algorithm

• Learning used four Sentinel-2, Sentinel-1, and Landsat inputs with reconstruction, batch-uniformity, input-consistency, and text-contrastive objectives.The reconstruction objective randomly holds out frames and uses source-specific, shift-invariant, re-gridding, and per-pixel weighted losses; text alignment used a frozen language model and CLIP loss.Source: S2.2.2. Learning algorithm · p. 27, S2.2.3. Reconstruction objective · p. 27, S2.2.6. Text-contrastive objective · p. 29

S2.2.4. Batch uniformity objective

• Batch uniformity used weight b=0.05 to discourage collapse, while a sweep found b=0.005 optimal, improving GLanCE land-cover BA from ~64.7% to ~66.6%.For LUCAS land-cover BA, b=0.005 yielded ~34.7% versus ~34.2% for b=0.Source: S2.2.4. Batch uniformity objective · p. 28

S2.2.5. Consistency objective

• The consistency objective sought embeddings that remain consistent despite irregular inputs, and with weight c=0.02, tile artifacts remained visible in the embedding fields.Source: S2.2.5. Consistency objective · pp. 28–29

S2.3. Model training; S2.4. Architectural details

• AEF trained for 56 hours on 512 TPU v4 devices over 100,000 steps, using 103-frame subsampled sequences, Adam optimization, 15 STP blocks, and fixed VMF concentration κ=8e3.The architecture used D_T=512 for time, D_S=1024 for space, and two-hidden-layer width-512 implicit decoders.Source: S2.3. Model training · p. 29, S2.4. Architectural details · p. 29

S3. Evaluation datasets

• Ten public reference datasets were selected for classification, regression, and change detection, prioritizing human interpretations or physical measurements, global coverage, precise point coordinates, and labels valid from 2017 onward.The suite generally avoids harmonized annotations with differing protocols and evaluations requiring shared source imagery, enabling comparisons across approaches using different inputs.Source: S3. Evaluation datasets · p. 30, S3.1. Selection criteria · p. 30

S3.2. Processing

• The standardized datasets produced 15 derivative evaluations, with spatial proximity filtering, class- or bin-balanced training splits, and remaining points assigned to test sets.Sampled points were separated by at least 1.28 km, and processed evaluation datasets were released as a supplemental dataset.Source: S3.2. Processing · p. 30

S3.3.1. LCMAP

• LCMAP provides four CONUS evaluations—land cover, land use, and their respective change tasks—using annual 2017–2021 labels to represent operational national-scale mapping and change detection.The final classification evaluations each use 1,800 training points, while change evaluations use 600 and 300 training points for land-cover and land-use change, respectively.Source: S3.3.1. LCMAP · pp. 32–33

S3.3.2. LUCAS

• LUCAS supplies detailed ground-based land-cover and land-use evaluations, with 40 and 15 classes respectively, using instantaneous validity for cover and a one-year window for land use.After processing and spatial filtering, the evaluations contain 12,000 training and 191,569 test points for land cover, and 4,500 training and 222,358 test points for land use.Source: S3.3.2. LUCAS · p. 34

S3.3.3. GLaNCE land cover

• GLaNCE contributes a global land-cover evaluation using Level 2 labels to test thematic detail, with random annual sampling across valid label segments to reduce temporal autocorrelation.The final evaluation contains eleven classes, 3,300 training points, and 31,585 test points.Source: S3.3.3. GLaNCE land cover · pp. 36–37

S3.3.4. Africa crop mask

• The Africa crop mask is a high-confidence binary cropland evaluation across eight Sub-Saharan African countries, based on multi-annotator agreement and annual country-specific valid periods.The final dataset contains 400 training points and 2,156 test points after preprocessing and spatial filtering.Source: S3.3.4. Africa crop mask · pp. 37–38

S3.3.5. Canada crops

• Canada crops provides coarse and fine national-scale crop classifications from sparse ground-based windshield surveys, using hierarchical labels and proportional 60/40 training-test splits.The final coarse evaluation has 2,831 training and 13,248 test points, while the fine evaluation has 5,565 training and 9,001 test points; fine categories were merged or removed when sparse or noisy.Source: S3.3.5. Canada crops · pp. 38–40

S3.3.6. Ethiopia crops

• EthCT2020 adds harmonized, quality-controlled, georeferenced in-situ annual crop-type samples from Ethiopian smallholder farming systems as another sparse-but-high-quality evaluation source.The dataset was downloaded from Mendeley and further processed for consistency with the evaluation protocols.Source: S3.3.6. Ethiopia crops · p. 40
• Ethiopia crops contains spatially filtered field plots labeled by crop group and class, yielding 873 training and 1,657 test points for the 2020/21 Meher season.Connected-component splitting joins points within 1.28 km, removes crop classes with fewer than 49 training points, and treats validity as instantaneous.Source: S3.3.6. Ethiopia crops · p. 40

S3.3.7. US trees

• US trees uses filtered GBIF/iNaturalist observations to evaluate genus-level mapping across 39 US tree genera, with 11,700 training and 33,682 test points.Records span January 1, 2017 to January 1, 2023, require at most 10 meters spatial uncertainty, and allocate 300 samples per class to training.Source: S3.3.7. US trees · pp. 40–41

S3.3.8. Descals oil palm

• Descals oil palm provides a three-class evaluation of non-plantation, industrial, and smallholder closed-canopy oil palm, with 600 training and 16,877 test points.Labels reflect plantation status during 2016–2021, and each row is randomly assigned a valid year from 2019 or 2021.Source: S3.3.8. Descals oil palm · pp. 41–42

S3.3.9. OpenET ensemble

• OpenET evaluates proxying the computationally intensive ensemble product rather than ET estimation directly, using 3,000 training and 32,683 test points.CONUS samples are stratified across monthly ensemble ET values from 0 to 200 mm, with entire-month validity and a preceding one-year support period.Source: S3.3.9. OpenET ensemble · pp. 42–43

S3.3.10. ASTER Global Emissivity Database (GED)

• ASTER-GED evaluates proxying a complex emissivity workflow using filtered 100-meter observations, with 2,000 training and 15,636 test points.The evaluation uses the 8.3 μm emissivity band, selects land locations within ±60° latitude, and treats 2017 as both support and valid periods despite the source product spanning 2000–2008.Source: S3.3.10. ASTER Global Emissivity… · pp. 43–44

S4. Details on evaluation setup

• Evaluation spans 15 datasets with trials varying label sparsity and transfer methods.Each transfer method is fit on training data and evaluated on held-out validation labels.Source: S4. Details on evaluation setup · p. 44, S4.1. Predictors · p. 44

S4.1. Predictors

• Predictors use low-parameter linear probes or kNN to support low-shot evaluations without undue hyperparameter penalties.Linear classification uses one-vs-rest ordinary least squares with λ = 0, while regression uses simple least squares with λ = 0; kNN uses l2 distance, majority voting, or averaging.Source: S4.1. Predictors · pp. 44–45
• Change classification concatenates before/after embeddings, whereas unsupervised anomaly detection thresholds normalized pairwise embedding differences to maximize balanced accuracy.The anomaly threshold is selected globally from 0.1 through 0.9, after which all other metrics are computed.Source: S4.1. Predictors · p. 45

S4.2. Kappa adjustment

• Kappa-adjusted metrics rescale scores relative to a random predictor and clamp error-rate scores exceeding random performance at 1.For Balanced Error Rate, a score of 1 is mapped to the balanced error rate of a random predictor.Source: S4.2. Kappa adjustment · p. 45

S4.3. Max trial group folds

• Max-trial folds are based on the least-represented training class, yielding k = 273 folds when that class has 75 labels and k = 1 when classes are balanced.This procedure addresses unequal numbers of labels across training classes.Source: S4.3. Max trial group folds · p. 45

S4.4. Uncertainty estimation

• Uncertainty uses class-balanced k-fold resampling for low-shot trials, 100 bootstrap resamples of validation data for full trials, and nested bootstrap procedures for smaller datasets.Small datasets combine multiple training sets with nested validation resampling when balanced sets cannot reach 200 examples per class, or 200 examples overall for regression.Source: S4.4. Uncertainty estimation · pp. 45–46

S5. Baseline comparisons

• Baseline comparisons span EO feature engineering, adapted geospatial deep-learning models, geographic controls, and a generic ImageNet vision model.The benchmark includes controls for purely geographic predictive power and for computer vision systems not designed specifically for Earth observation.Source: S5. Baseline comparisons · p. 46
• Baseline features are standardized by resampling coarse outputs to 10m and extracting values at precise evaluation locations, while processing must be repeated for each temporal window or spatial extent.Sub-pixel aliasing is retained because label coordinates are stored at full precision.Source: S5. Baseline comparisons · p. 46

S5.1.1. XY / S5.1.2. XYZ

• XY and XYZ provide four- and five-dimensional positional controls, but neither varies through time and both are expected to perform poorly on dynamic landscapes or change detection.XYZ augments trigonometric longitude and latitude with normalized elevation from the Copernicus GLO-30 DEM.Source: S5.1.1. XY · pp. 46–47, S5.1.2. XYZ · p. 47

S5.1.3. Vision Transformer (ViT)

• The ViT control uses ImageNet-pretrained RGB Sentinel-2 inputs, averages masked outputs across time into a 1024-dimensional embedding, and was tuned to maximize evaluation performance.Its design is constrained by standard ViT handling of resolution, spectral bands, and multitemporal imagery; the selected variant outperformed non-controls in some instances.Source: S5.1.3. Vision Transformer (ViT) · pp. 47–48

S5.2.1. Composites / S5.2.2. CCDC / S5.2.3. MOSAIKS

• Composite and CCDC baselines represent spectral and temporal information using 16- and 56-dimensional feature vectors, respectively, while MOSAIKS provides a spatially localized nonlinear representation.Composites use medians for optical sources and means for radar. CCDC selects eight harmonic coefficients plus RMSE values for seven Landsat bands.Source: S5.2.1. Composites · p. 48, S5.2.2. Continuous Change Detection… · p. 49, S5.2.3. MOSAIKS · p. 49

S5.3.1. SatCLIP / S5.3.2. Prithvi / S5.3.3. Clay

• SatCLIP is evaluated with only its location encoder on longitude and latitude, whereas Prithvi is a temporal pretrained architecture whose embeddings use 1–3 frames.SatCLIP embeddings have 256 dimensions and are not time-varying; Prithvi embeddings have 768–2304 dimensions.Source: S5.3.1. SatCLIP · p. 50, S5.3.2. Prithvi · p. 50

S6.1. Comparisons with 10 and 1 samples per-class

• 500-fold 10-shot trials exceeded 1.0x error reduction within the ~90% confidence interval in only 8/15 evaluations, versus 5/15 for 1000-fold 1-shot trials.Average variation was ±0.38x for 10-shot trials and ±0.49x for 1-shot trials, indicating that reliable general 1-shot and 10-shot performance remains unresolved despite mean gains favoring AEF.Source: S6.1. Comparisons with 10 and 1… · p. 51

S6.2. Classification

• AEF consistently led classification evaluations, enabling strong low-shot performance across binary, land-cover, land-use, and tree-species mapping tasks.The next-best approach varied substantially, and some baselines performed no better than the null expectation; AEF generally achieved the highest balanced accuracies, except on Ethiopia crops, where kNN with k=1 was preferred.Source: S6.2. Classification · pp. 51, 53, Figure S16 · p. 52
• Task-specific baselines were competitive in distinct settings: CCDC for crop mapping, SatCLIP for Ethiopia crops and US trees, and MOSAIKS or Clay for land-cover and land-use mapping.These patterns suggest phenology, localized EO information, or spatial context can be especially valuable depending on the application.Source: S6.2. Classification · p. 51

S6.3. Regression

• AEF achieved the best overall regression performance in both R2 gains and MAE reductions, remaining viable across all tested OpenET predictor combinations.Other methods produced negative R2 values for ASTER GED or OpenET; only composites with linear probes and MOSAIKS with kNN k=3 besides AEF achieved R2 > 0 on OpenET.Source: S6.3. Regression · p. 53, Figure S17 · p. 53, Figure S18 · p. 53
• For ASTER emissivity, MOSAIKS was next-best and composites performed strongly, whereas CCDC and positional encodings performed worse, implicating local spatial patterns and overall reflectance.For OpenET, AEF was the only approach with viable results across all tested kNN and linear predictors.Source: S6.3. Regression · p. 53

S6.4. Change detection

• Change-detection performance was generally less differentiated, with Prithvi and linear MOSAIKS trials near the random baseline after location-only baselines were excluded.SatCLIP, XY, and XYZ were omitted because they lack temporal handling or cannot distinguish observations across time.Source: S6.4. Change detection · p. 54, Figure S19 · p. 54

S7.1. Training observations

• AEF generally improved monotonically as observations increased for 9-of-15 evaluation datasets, despite unpredictable nonmonotonicity elsewhere and no obvious grouping or regional bias.Baselines showed no obvious common trend, indicating that methods performed better on different problems.Source: S7.1. Training observations · p. 54

S7.2. Sources

• All source groups produced the best performance in 11-of-15 evaluations, although different tasks preferred different measurement combinations.The Descals oil palm evaluation favored Optical + Radar + LiDAR, while all non-change LULC evaluations favored including Annotated sources.Source: S7.2. Sources · p. 54, Figure S21 · p. 56

S7.3. Bottleneck characteristics

• The reconstruction bottleneck trades information capacity against latent-manifold smoothness through embedding dimension and VMF κ, with AEF using Embedding D=64 and κ=8e3.Lower κ increases smoothness through noise regularization, whereas higher κ increases channel and embedding capacity; the selected setting was evaluated across transfer methods and trial sizes.Source: S7.3. Bottleneck characteristics · pp. 54, 61, Figure S22 · p. 57
• Bottleneck performance varied considerably by evaluation, with larger-legend datasets generally favoring more concentrated noise and higher embedding dimensions in the max-trial setting.This trend appeared for Canada crops (fine), LUCAS-derived datasets, and iNaturalist trees; noisier bottlenecks improved or did not affect smaller trial groups, especially Canada crops and OpenET ensemble.Source: S7.3. Bottleneck characteristics · p. 61

S8.1. Quantization

• Quantization transforms 32-bit floats into signed 8-bit or 16-bit integers, applying power-law scaling, rounding, clipping, and dequantization.The exponentiation was introduced to preserve information in the least significant digits of dequantized values.Source: S8.1. Quantization · p. 61

S8.2. Creating embedding fields

• Global annual embedding fields are produced by buffered 960m x 960m UTM-zone tiles, trimming outer 80m before rendering to avoid seams across projections.The Earth Engine-backed system scales to hundreds of billions of observations while using the same source-collection protocol as training.Source: S8.2. Creating embedding fields · p. 62

S8.3. Production model (AEF v2.1)

• AEF v2.1 expands training coverage, corrects crop-classification regression and processing bugs, uses full-year inference imagery, and mitigates tiling and re-gridding artifacts.Training video sequences increased from 8,412,511 to 10,182,450 after removing target and sensor-presence requirements.Source: S8.3. Production model (AEF v2.1) · p. 62

S8.1. Quantization

• 8-bit quantization with power = 2 showed little performance variability relative to non-quantized embeddings and offered the best storage/performance tradeoff.The quantization exponent preserves information in least significant dequantized digits; Figure S23 reports regression and change-detection results.Source: S8.1. Quantization · p. 61, Figure S23 · p. 63