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

Priyanka Goyal; Maitreyee Sar; Adway Mitra; Manjira Sinha

cs.CVOct 27, 2026

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

Authors: Priyanka Goyal, Maitreyee Sar, Adway Mitra, Manjira Sinha

Organizations: Department of Artificial Intelligence, Indian Institute of Technology Kharagpur, West Bengal, India · Centre for Atmospheric Sciences, Indian Institute of Technology Delhi, India · Tata Consultancy Services Research, Kolkata, India

Abstract

Earth Observation (EO)-based agricultural monitoring is shifting from pixel-level spectral analysis toward representation-driven modeling, yet systematic evaluation of learned representations across spatial, temporal, and domain-generalization dimensions remains limited. Geospatial foundation models generate high-dimensional embeddings encoding spatio-temporal land-surface properties, but their utility for agricultural applications remains insufficiently understood. This study presents a representation-agnostic evaluation framework integrating geometric separability, temporal dynamics, spatial robustness, and cross-regional generalisation. A 64-dimensional AlphaEarth embedding space is benchmarked against Sentinel-2 reflectance, NDVI-derived phenology, and harmonic features using field-scale data from the Conterminous United States (CONUS; 2017-2023) and Northern India (2021). Traditional features achieve strong linear separability (Fisher ratio), while embeddings provide superior distribution-aware separation -- Bhattacharyya and Mahalanobis distances up to an order of magnitude higher for spectrally similar pairs (e.g. Corn/Soybeans: 3.9 vs. 0.10). Embeddings uniquely capture multi-year crop rotation dynamics, reaching a separability ratio of 1.784 versus a maximum of 1.003 across all engineered features, and exhibit lower residual spatial autocorrelation of errors (mean Moran's $I \approx 0.17$ ). However, harmonic and NDVI features show stronger cross-continental transferability (US $\to$ India Macro F1: 0.493 vs. 0.326 for embeddings), reflecting better distributional alignment. The combined feature space yields balanced performance, supporting hybrid modeling. These findings reveal a fundamental trade-off between representational richness and cross-region transferability, providing a structured framework for feature selection in large-scale crop classification.

Figures & tables

Figure 1: Overview of the representation evaluation framework. Feature sets (latent embeddings, spectral reflectance, NDVI phenology, harmonic features) are extracted at field level from Google Earth Engine and evaluated across four dimensions: (i) geometric separability, (ii) temporal dynamics and rotation structure, (iii) spatially explicit validation, and (iv) cross-regional transfer analysis for both CONUS and Northern India datasets.

Figure 2: Study regions across five U.S. counties (Fresno, CA; Finney, KS; Lancaster, NE; McLean, IL; Bolivar, MS). Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure 3: Frequency distribution of crop rotation types across the five CONUS study counties, including monoculture (single-crop), alternating (two-crop, predominantly corn–soybean), three-crop, and multi-crop (

\geq

4 crops) systems. The diversity of rotation regimes provides a rigorous testbed for evaluating multi-year temporal feature representations.

County	GEOID	Fields	Crops	Mean Area (ha)	Area Range (ha)
Fresno	06019	1,748	21	128.2	10.1–11,252.0
McLean	17113	929	10	259.7	5.0–19,449.9
Finney	20055	1,759	16	103.2	5.0–8,130.3
Bolivar	28011	580	9	76.4	5.0–2,090.4
Lancaster	31109	1,299	9	114.8	5.0–5,919.6

Table 1: Summary of the 2017–2023 CONUS dataset. GEOID = US Census Bureau county FIPS code (Federal Information Processing Standard geographic identifier). “Crops” indicates the number of unique crop classes observed across the CDL time series for each county.

Overall		Crop distribution		State distribution		Field area
Total fields	3,890	Wheat	52.0%	Uttar Pradesh	94.0%	Mean (ha)	0.256
States covered	4	Mustard	25.3%	Odisha	3.1%	Std. dev.	0.331
Crop classes	12	Maize	7.5%	Madhya Pradesh	2.9%	Median (ha)	0.148
		Sugarcane	4.2%			$<0.1$ ha	32.7%
		Rice	3.3%			$0.1$ – $0.5$ ha	54.8%
		Lentil	2.6%			$0.5$ – $1.0$ ha	8.7%

Table 2: Summary of the AgriFieldNet dataset (India, 2021).

Evaluation Criterion	Embedding	Spectral	NDVI	Harmonic	Combined
Geometric Separability
Linear separability (Fisher ratio)	+	+	+	+	++
Distribution-aware separation (Bhattacharyya)	++	-	-	-	++
Distribution-aware separation (Mahalanobis)	++	-	-	-	++
Feature space interpretability	-	+	++	+	-
Intrinsic Structure

Table 3: Suitability of feature representations across evaluation criteria. Ratings derived from quantitative results in Sections 3.2 – 3.9 . ++ = strongly suitable; + = suitable;

\sim

= moderate; - = limited. Combined: combined embedding + engineered feature set.

Figure 4: Embedding space structure across feature representations. (a) Cross-correlation between the 64 embedding dimensions and each engineered feature set. Near-zero means (

|\mu|<0.03

) confirm no systematic global alignment; wider spread for NDVI (

\sigma=0.22

) reveals partial phenological overlap. (b) Effective dimensionality:

n_{95}

(bars) and participation ratio (PR; markers) for the US (solid) and India (hatched) datasets. Embeddings occupy a substantially higher-dimensional regime (

n_{95}>20

\text{PR}>7.5

) compared with all engineered feature sets (

n_{95}\leq 3

\text{PR}\leq 2.7

). (c) Cumulative variance explained as a function of retained principal components (US dataset). Engineered features concentrate 95% of variance in two to three components, whereas embedding variance is spread across more than 20 components.

Figure 5: Feature importance across US (red bars) and India (blue bars) datasets, ranked by PCA loading magnitude (top row) and ANOVA

F

-statistic (bottom row). For embeddings, the top PCA dimensions (A38, A60) differ markedly from the top ANOVA dimensions (A56, A49) for both US (red) and India (blue), confirming variance–discriminability decoupling : the directions of greatest variance are not the directions of greatest class discrimination. NDVI exhibits the most consistent feature hierarchy between PCA and ANOVA rankings (Spearman

\rho=0.40

–

0.80

), implying that variance-based compression retains most discriminative information for this feature set. For harmonic features, the ANOVA ranking is anti-correlated with PCA (

\rho=-1.00

for the US, red): harmonic amplitude dominates PCA but is the least discriminative feature, while phase angle—stable across years but crop-specific—carries the most class-relevant information.

Figure 6: ANOVA $F$ -statistics versus PCA loading magnitudes for individual embedding dimensions (red circles = dimensions with the highest ANOVA $F$ -statistic; blue circles = dimensions with the highest PC1 loading magnitude). Several high-ANOVA dimensions (A56, A49; shown in red) lie far from the dominant PCA directions (A38, A60; shown in blue), confirming that class-relevant discriminative information is distributed across non-principal axes. PCA-based dimensionality reduction would therefore discard approximately half of the class-relevant embedding information, since dimensions that explain the most variance (high $|$ PC1 loading $|$ ) are not the same as those that best discriminate between crop classes (high ANOVA $F$ ).

Feature Set	Macro F1	Weighted F1	Macro F1	Weighted F1	L1	ENet †	$\rho$ (PCA,LR)	$\rho$ (ANOVA,LR)
	Linear Sep. (LR-L2)		Non-linear Sep. (MLP)		Sparsity (retained dims)		Ranking Agreement
Embedding	$\mathbf{0.09\pm 0.07}$	$\mathbf{0.27\pm 0.16}$	$\mathbf{0.065\pm 0.053}$	$\mathbf{0.225\pm 0.171}$	64/64	64/64	0.14	0.29
Spectral	$0.05\pm 0.03$	$0.20\pm 0.13$	$0.041\pm 0.027$	$0.196\pm 0.142$	6/6	6/6	0.26	0.09
NDVI	$0.05\pm 0.03$	$0.22\pm 0.17$	$0.047\pm 0.025$	$0.240\pm 0.185$	4/4	4/4	$\mathbf{0.40}$	$\mathbf{0.80}$
Harmonic	$0.05\pm 0.03$	$0.22\pm 0.16$	$0.035\pm 0.025$	$0.212\pm 0.178$	5/5	5/5	0.30	-1.00

Table 4: Linear and non-linear separability, feature sparsity, and feature ranking agreement across feature sets (US data, LORO-CV). Low overall F1 scores reflect the difficulty of cross-county generalisation under Leave-One-Region-Out validation; differences across feature sets highlight representational properties rather than absolute classification performance.

Figure 7: Class separability across six complementary metrics for US (blue-shaded) and India (red-shaded) crop pairs. Crop pair abbreviations on the x-axis: C/Sor = Corn vs. Sorghum; C/Soy = Corn vs. Soybean; C/Wht = Corn vs. Wheat; S/Wht = Soybean vs. Wheat; R/Wht = Rice vs. Wheat (US and India); W/Mz = Wheat vs. Maize (India); Mz/R = Maize vs. Rice (India). US pairs (blue background, left of dashed line); India pairs (red background, right of dashed line). (a) Fisher ratio: spectral and harmonic features achieve competitive mean-based separation for several US pairs, but all feature sets converge near zero for Corn/Soybean, reflecting their spectral similarity. (b–c) Bhattacharyya and Mahalanobis distances: embeddings consistently outperform all engineered features by up to an order of magnitude (C/Soy Bhattacharyya: 3.9 embeddings vs. 0.10 spectral; S/Wht Mahalanobis: 18.3 vs. 4.1 NDVI; R/Wht India: 57.5 vs. 5.8 NDVI). (d) Jeffries–Matusita distance saturates at the theoretical maximum of 1.41 for embeddings across all US pairs, confirming complete distributional separation. (e–f) AUC (Area Under the ROC Curve) and log-loss corroborate improved probabilistic classification for embeddings; error bars represent standard deviation across five cross-validation folds.

Figure 8: Temporal dynamics and crop rotation structure across feature representations. (a) Multi-year rotation trajectories projected onto the first two principal components. Embedding-based trajectories form coherent, rotation-type-specific loops consistent across all five counties, whereas spectral and NDVI feature trajectories collapse into an undifferentiated cloud. (b) Within-class temporal stability (SDI; dimensionless, range 0–2, log scale). Lower SDI indicates that a field’s trajectory closely tracks its class centroid trajectory. Embeddings occupy an intermediate regime (

\text{SDI}\approx 0.277

) balancing sensitivity to annual crop dynamics with resistance to year-to-year noise. (c) Rotation-type separability ratio (bars) and mean ANOVA

F

-statistic (diamonds). Only embeddings surpass the critical threshold of 1.0 (ratio = 1.784; all 64 dimensions significant at

p<0.001

Feature Set	Within var.	Between dist.	Sep. ratio	Mean F	Sig. dims
Embedding	0.178	0.753	1.784	450.67	64/64
Spectral	0.006	0.074	1.003	127.07	6/6
NDVI	378.04	18.21	0.936	348.28	4/4
Harmonic	1.136	1.055	0.990	161.29	5/5
Combined	379.35	18.33	0.941	402.59	79/79

Table 5: Rotation type separability by feature set. Separation ratio = mean between-rotation distance /

\sqrt{\text{mean within-rotation variance}}

. Mean F = mean ANOVA

F

-statistic across all feature dimensions (higher = greater discriminative power per dimension). Sig. dims = number of dimensions with statistically significant between-rotation discrimination (

p<0.001

after permutation testing), expressed as a fraction of the total feature dimensionality. All permutation tests use 1,000 iterations.

Figure 9: Spatial validation: classification performance and spatial error structure. (a) Model-averaged Macro F1 across Tier 1 LORO validation and three Tier 2 spatial block cross-validation sites. Embedding-row cells are outlined in purple. (b) Individual classifier performance under LORO-CV. The TCN achieves the highest overall performance (Macro F1 = 0.50) with combined features. (c) Macro F1 versus Moran’s $I$ on classification errors. Embedding and combined features cluster in the upper-left ideal zone (high F1, low spatial error autocorrelation). Harmonic features exhibit high Moran’s $I$ ( $\approx$ 0.4 at Fresno), indicating spatially structured errors that inflate apparent accuracy under non-spatial cross-validation.

Figure 10: Cross-regional consistency: structural alignment (CKA) versus distributional shift (EMD). Abbreviations in panel labels: MW = Midwest group (McLean, IL; Finney, KS; Lancaster, NE); P = Plains group (Finney, KS; Lancaster, NE). (a) CKA–EMD trade-off scatter for all pairwise within-US county comparisons (filled markers; e.g. Fresno vs. MW/P) and US

\to

India cross-continental comparisons (open markers). The green-shaded quadrant (low EMD, moderate-to-high CKA) represents the preferred transferability region. (b) CKA (bars; left axis) and EMD (hatched bars; right axis) for the US

\to

India distributional shift by feature set. Harmonic features maintain the lowest distributional shift (lowest EMD); embeddings achieve the highest structural alignment (highest CKA). (c) Normalised CKA values across spatial scales, from within-US pairwise comparisons to cross-continental US–India transfer. All feature sets experience substantial CKA collapse (

\leq 0.20

of within-US maximum), confirming the severity of the cross-continental domain shift.

Figure 11: Classification performance across feature representations: within-region and cross-continental. (a) Per-crop F1 scores under within-region evaluation for the US (filled markers) and India (open markers). (b) Cross-continental Macro F1 for both transfer directions. Harmonic features achieve the highest cross-continental Macro F1 (US

\to

India: 0.493; India

\to

US: 0.389), substantially outperforming embeddings (0.326; 0.198). (c) Mean F1 across crops as a function of evaluation context. Harmonic and NDVI features show shallower performance slopes, confirming superior transferability.

Figure 12: Classification performance (Weighted F1) as a function of field area across India (

<

0.5 ha) and U.S. (

>

5 ha) datasets. Embedding and combined features maintain robust performance across the full size range (

\geq

0.75 at

>

0.2 ha), whereas NDVI-based models exhibit significant performance drops in small fields (

<

0.1 ha,

\Delta

\approx

0.30). The orange-shaded region marks India-dominated bins; the blue-shaded region marks U.S.-dominated bins.

Agricultural Image Analysis

Agricultural image analysis uses computer vision and multimodal learning to monitor crops, livestock, produce quality, and field conditions.

See all 94 papers

Sep 15, 2026

From Foundation Embeddings to Cropland Maps: Label Efficiency, Temporal Transferability and Independent Human Validation

Geospatial foundation models provide reusable representations of satellite imagery that support downstream mapping with limited task-specific modelling. We evaluate whether annual AlphaEarth embeddings support binary cultivated-versus-non-cultivated mapping in Maine, USA, using 192 spatially separated patches and labels derived from the USDA Cropland Data Layer (CDL). Without fine-tuning the foundation model, a lightweight classifier reaches 93.7% overall accuracy and 90.8% balanced accuracy on held-out patches. Logistic regression is within 0.3 percentage points of a gradient-boosted ensemble, while a nearest-class-centroid rule, which uses class centroids but fits no parameters, reaches 90.2%. A balanced sample of 60,000 labelled pixels is within 1.3 percentage points of the full pool of 8.6 million pixels; because pixels are spatially autocorrelated, this result concerns pixel-sample efficiency rather than 60,000 independent annotation sites. In a same-region transfer experiment, classifiers trained in one year remain accurate across 2018 to 2023. Against a blind, two-interpreter consensus at 385 randomly sampled points in one contiguous 2023 block, the AlphaEarth-plus-random-forest map agrees at 95.3% (

κ=0.82

), compared with 91.7% for the CDL (

κ=0.72

; exact two-sided McNemar

p=0.0161

). This local result is consistent with partial smoothing of CDL label noise, but it does not establish statewide correction of the reference product. On the same points, the difference from a fine-tuned TerraMind segmentation model is not statistically significant (95.3% versus 93.5%;

p=0.14

), and the experiment is not a controlled comparison of computational cost. These results support frozen geospatial embeddings as a low-compute candidate for regional cropland mapping, subject to the limits of a single-state study, a 30 m-derived training reference, and a one-block human validation.

Mohammad Ammar Mughees, Giovanni Montefoschi, Zhongxin Chen +1

Department of Civil and Environmental Engineering, Politecnico di Milano, Milan, Italy · Food and Agriculture Organization of the United Nations, Rome, Italy

Agricultural Remote Sensing Geospatial Foundation Models Geospatial Representation Learning Land Cover Classification Transfer Learning

Ask PDF

May 20, 2026

Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis

Field-scale crop maps support supply-chain forecasting and policy, yet statewide crop identification still often depends on retrospective surveys or remote-sensing workflows built around hand-engineered spectral features. Those pipelines can be accurate, but they require repeated preprocessing and often lose robustness across years. This study evaluated whether Google DeepMind's AlphaEarth geospatial embeddings can serve as an analysis-ready alternative for mapping processing tomato systems in California. LandIQ 2018 crop polygons were used to assemble a balanced reference dataset of 4,742 tomato and 4,742 non-tomato fields. For each polygon, 64-band AlphaEarth embedding chips were extracted and aligned with binary masks, then divided into spatially independent training (n = 6,638), validation (n = 1,422), and test (n = 1,424) sets. A U-Net segmentation model was trained on AWS SageMaker using a composite masked binary cross-entropy and soft Dice loss. To complement hard predictions, Monte Carlo dropout was retained at inference and repeated 100 times per chip to estimate predictive mean and variance. On the independent test set, the model achieved 99.19% pixel accuracy, 98.69% precision, 99.40% recall, 99.04% F1 score, 98.11% intersection over union, and 99.02% chip accuracy. Uncertainty maps were consistently highest near field edges and low within field interiors. The results show that AlphaEarth embeddings retain crop-relevant spatial and temporal structure and can support accurate, field-scale tomato mapping without manual feature engineering.

Mohammadreza Narimani, Alireza Pourreza, Parastoo Farajpoor

Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, 95616, USA

Remote Sensing Image Segmentation Agricultural Remote Sensing Geospatial Foundation Models

Ask PDF

Oct 27, 2026

Ground-to-Satellite Knowledge Distillation for Crop Classification under Temporal and Geographic Shift

Continuous and up-to-date crop type mapping is essential for agricultural monitoring but is often hampered by limited reference data. The growing availability of ground-level field photographs offers a promising way to address this challenge. When combined with satellite time series, these images provide complementary perspectives on the same parcel, capturing crop characteristics from distinctly different perspectives. However, field photographs are not always available for each satellite pixel at inference time, limiting their potential in scalable crop type mapping. We investigate whether ground photographs can provide privileged supervision for crop classification built on pretrained satellite embeddings. A photo teacher based on frozen CLIP features supervises classifiers trained on AlphaEarth and TESSERA representations; photographs are used only during training. On a paired field photo and satellite imagery dataset with 17 crop classes, knowledge distillation consistently improves over satellite-only classification; for frozen embeddings the margin grows under temporal shift, reaching up to +5.5 macro-F1 points, and widens in geographic holdouts (+3.7 vs +3.0 in-distribution). These results show that ground-view supervision can complement satellite embeddings and improve their robustness without changing inference requirements.

Miro Miranda, Rocco Sedona, Valerio Marsocci +2

Jülich Supercomputing Centre, Forschungszentrum Jülich, Germany · European Space Agency (ESA/ESRIN), Italy · University of Iceland, Iceland +1

Agricultural Remote Sensing Knowledge Distillation Image Classification Distribution Shift

Ask PDF

Abstract

Agricultural Image Analysis

Explore similar work

From Foundation Embeddings to Cropland Maps: Label Efficiency, Temporal Transferability and Independent Human Validation

Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis

Ground-to-Satellite Knowledge Distillation for Crop Classification under Temporal and Geographic Shift