Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability

• A representation-agnostic framework benchmarks AlphaEarth embeddings for crop classification.Source: Highlights · p. 2
• Embeddings provide up to 39 times stronger distribution-aware class separability than spectral features.Source: Highlights · p. 2
• Only embeddings capture multi-year crop rotation dynamics, with a separability ratio of 1.784 versus ≤1.003 for engineered features.Source: Highlights · p. 2
• Embedding classifiers show lower spatial error clustering, with mean Moran’s I ≈0.17.Source: Highlights · p. 2
• Harmonic features transfer better across continents, while hybrid features balance transferability and representation quality.Source: Highlights · p. 2

• Crop datasets and conventional engineered features provide limited insight into multi-year rotations and management-driven variability, despite their use in agricultural monitoring.NDVI-based features also face saturation, atmospheric sensitivity, and limited capacity to represent complex variability.Source: §1 · p. 2
• Temporal trajectories can distinguish spectrally similar crops, but periodic time-series methods may miss irregular rotations and management variability.Source: §1 · p. 2
• Evaluation of remote-sensing representations should extend beyond predictive accuracy to include structural and statistical properties, since greater capacity does not ensure robustness under spatial or domain shifts.Source: §1 · p. 2
• Whether learned embeddings capture crop phenology, rotations, and spatially consistent patterns—or mainly statistical regularities—remains unclear under domain shift.Source: §1 · p. 3
• The study introduces an architecture-agnostic framework spanning separability, temporal trajectories, spatial validation, and cross-regional analysis in contrasting CONUS and Northern India farming systems.Source: §1 · p. 3

RQ4: Do embeddings generalize better across distinct agro-ecological regions?

• The evaluation framework jointly analyzes linear and distribution-aware separability to characterize feature-space geometry.Source: §4 · p. 4
• A trajectory-based approach models multi-year crop dynamics, including rotation patterns and management-driven signals.Source: §4 · p. 4
• Spatially explicit validation combines LORO and spatial block cross-validation with Moran’s I analysis of residual errors.Source: §4 · p. 4
• Cross-regional consistency is assessed with Centered Kernel Alignment and Earth Mover’s Distance to characterize structural and distributional alignment.Source: §4 · p. 4
• The study frames feature selection as a trade-off between representation quality and cross-regional transferability.Source: §4 · p. 4

2.1. Study Areas, Data Processing, and Feature Construction

• Four datasets are resampled to a common 10 m resolution, temporally filtered to crop-relevant growing seasons (DOY 90–300 for CONUS and region-specific windows for India), and aggregated from pixels to field units.CONUS fields require stable CDL labels across at least six years; India processing applies a 5 m interior buffer.Source: §2.1 · pp. 4, 6
• Four field-level feature sets comprise 64-dimensional latent embeddings, Sentinel-2 reflectance, NDVI phenology summaries, and one-cycle harmonic features.Features are constructed using mean aggregation; reflectance uses median growing-season values, while NDVI features include extrema, amplitude, and peak timing.Source: §2.1 · p. 6

2.2. Representation Evaluation Framework

• PCA and maximum-likelihood estimation characterize intrinsic structure, while Spearman correlation compares feature-importance rankings from PCA loadings, ANOVA F-statistics, and model-based importance.The framework also uses eigenvalue decay and participation ratio to characterize representational structure.Source: §2.2 · pp. 7–8
• Separability is assessed with Fisher, Bhattacharyya, Mahalanobis, and Jeffries–Matusita metrics, plus L2-regularized logistic regression and a single-hidden-layer MLP.All features are standardized using z-score normalization.Source: §2.2 · p. 8
• Multi-year field trajectories are used to assess rotation separability by a between-class-distance to within-class-variance ratio, and temporal consistency by field-level SDI and class-centroid similarity across years.The CONUS observation window uses seven years, and lower SDI indicates greater temporal stability.Source: §2.2 · p. 8
• Spatial robustness uses Leave-One-Region-Out validation and 10 km × 10 km spatial block cross-validation, with Moran’s I applied to classification errors; CKA and EMD assess cross-regional structural and distributional alignment.The framework evaluates Random Forest, XGBoost, LightGBM, Support Vector Machines, LSTM, and Temporal Convolutional Networks.Source: Cross-Regional Consistency · p. 9

2.3. Cross-Continental Transferability and Scale Effects

• Cross-continental transfer trains on CONUS and tests on India, and vice versa, using maize, rice, and wheat; classification performance is also compared across field-size bins.Source: §2.3 · p. 9

3.2. Representation Structure

• Embeddings have higher effective dimensionality than engineered features, but PCA-based selection would discard approximately half of their class-relevant dimensions.Embedding n95 is 21 in the US and 30 in India, versus n95 ≤3 for engineered features; harmonic amplitude explains the most variance while phase angle is the primary discriminant.Source: §3.2 · pp. 9–10, Figure 4 · p. 11, Figure 5 · p. 12

3.3. Linear and Non-linear Separability

• Under strict spatial generalization, all feature sets have Macro F1 below 0.10, though embeddings score highest with linear models and nonlinear MLP models do not improve performance.All 64 embedding dimensions are retained under L1 and Elastic Net regularization.Source: §3.3 · p. 11, Table 4 · p. 11, §3.2 · p. 11

3.4. Distribution-Aware Separability

• For Corn/Soybeans, embedding Bhattacharyya distance is 3.9 versus 0.10 for spectral features, while Fisher ratios are virtually identical across feature sets at approximately 0.05.The Jeffries–Matusita distance reaches its 1.41 theoretical maximum for embeddings across US crop pairs.Source: §3.4 · p. 12, Figure 7 · p. 14

3.5. Temporal Dynamics and Crop Rotation Structure

• Embeddings achieved the highest crop-rotation separation ratio (1.784); engineered features ranged from 0.936 to 1.003, with Spectral at 1.003, Harmonic at 0.990, NDVI at 0.936, and Combined at 0.941.Embeddings’ intermediate SDI (≈0.277) balances sensitivity to annual crop dynamics with resistance to year-to-year noise.Source: §3.5 · p. 12, Figure 8 · p. 15

3.6. Spatial Robustness and Error Structure

• Under Tier 1 LORO validation, embeddings reach model-averaged Macro F1 0.34 versus 0.22–0.27 for engineered features, while embedding and combined-feature errors have mean Moran’s I ≈0.17–0.20.The corresponding Moran’s I for harmonic features is ≈0.43–0.48; the TCN reaches peak performance with combined features (Cohen’s κ = 0.57).Source: §3.6 · p. 13, Figure 9 · p. 16

3.9. Field Size Effects

• Embedding and combined-feature models maintain Weighted F1 ≥0.75 for fields larger than 0.2 ha, while NDVI-based models drop by ∆F1 ≈0.30 in fields smaller than 0.1 ha.Within-region performance is context-dependent: the combined feature set has the highest F1 in both regions, while spectral features outperform embeddings for several Indian crops.Source: §3.9 · pp. 15, 17, §3.8 · p. 15, Figure 12 · p. 17

3.7. Cross-Regional Consistency

• Harmonic features achieve cross-continental Macro F1 of 0.493 US→India and 0.389 India→US, exceeding embeddings at 0.326 and 0.198, respectively.Embeddings have higher structural alignment but greater cross-continental distribution shift; CKA collapses substantially for every feature set.Source: §3.7 · p. 14, §3.8 · p. 15, Figure 10 · p. 16

3.10. Implications for Agricultural Monitoring

• Embeddings suit within-region rotation monitoring and small-field classification, traditional spectral and phenological features transfer better across regions, and combined features generally balance these goals.The authors identify distributional mismatch as a concern for embedding transfer and point to feature selection by task and deployment context.Source: §3.1 · p. 9, §3.10 · pp. 17–18

4. Discussion

• Embeddings outperform engineered features in distribution-aware separability and rotation structure, but harmonic and NDVI features retain more within-region performance under cross-continental transfer.The discussion reports embedding transfer gap −0.50 and 40–50% retained performance for harmonic and NDVI features.Source: §4 · p. 18
• The embeddings were evaluated as fixed pre-trained representations without region-specific fine-tuning, so the reported results are described as a conservative lower bound on their utility.The authors note that lightweight target-domain adaptation has been shown to yield substantial performance gains.Source: §4 · p. 18
• The Indian validation data cover one growing season, 2021, and are concentrated in Uttar Pradesh, limiting generalisation across broader agro-ecological and calendrical contexts.94% of Indian fields are in Uttar Pradesh.Source: §4 · p. 19
• The authors identify domain adaptation, region-specific fine-tuning, multimodal fusion, and mapping embedding dimensions to biophysical variables as priorities for future work.They also note that a single-cycle harmonic fit may inadequately capture double-cropping systems.Source: §4 · p. 19

5. Conclusion

• The evaluation benchmarks 10 m AlphaEarth embeddings across separability, temporal dynamics, spatial validation, and cross-continental transfer using five CONUS counties and Northern India data.The authors state that the framework applies to any geospatial foundation model embedding space and provides a reusable community benchmark.Source: §5 · p. 19
• Bhattacharyya distances for Corn/Soybeans are 3.9 for embeddings versus 0.10 for engineered features, while Rice/Wheat India has an embedding Mahalanobis distance of 57.5.Embedding feature space also shows variance–discriminability decoupling, favoring discriminative over variance-based dimensionality-reduction criteria.Source: §5 · p. 19
• The embedding separability ratio is 1.784 for multi-year crop rotations, versus ≤1.003 for all engineered features, and mean residual-error Moran’s I is ≈0.17.The study reports embeddings as the only feature set capturing multi-year crop-rotation dynamics and the lowest residual spatial autocorrelation of errors.Source: §5 · p. 19
• Embedding performance degrades most under cross-continental transfer (gap = −0.50), while harmonic features retain 40–50% of within-US performance.The authors identify harmonic features as preferable for scalable global deployment.Source: §5 · p. 19
• The authors characterize embeddings as suited to within-region classification, rotation monitoring, and smallholder mapping, with fine-tuning, domain adaptation, and multimodal fusion identified as steps toward global monitoring.These approaches are presented as ways to address the cross-region transfer limitation.Source: §5 · p. 19

Data Availability

• AlphaEarth embeddings are available through Google Earth Engine and Google Cloud Storage, AgriFieldNet through Radiant Earth MLHub, and CDL through USDA NASS CropScape.Sentinel-2 Level-2A imagery is available via Google Earth Engine.Source: Data Availability · p. 20
• Code and derived datasets are planned for release through a public repository on GitHub and Zenodo upon acceptance.Source: Data Availability · p. 20
• The study describes its data in Section 2 and identifies public access routes for its listed datasets and imagery.Source: Data Availability · p. 20