PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Mohammad Ammar Mughees, Giovanni Montefoschi, Zhongxin Chen, Maria Antonia Brovelli

arXiv:2609.17138v1cs.CVcs.LGeess.IV

TL;DR

This paper asks whether reusable AlphaEarth satellite-image embeddings can support accurate cropland mapping with limited task-specific modelling and labels. Using frozen annual embeddings and USDA Cropland Data Layer-derived binary labels in Maine, it finds strong held-out accuracy, label efficiency, same-region temporal transfer, and closer agreement with independent human interpretation than the CDL.

  • Problem

    The paper asks whether geospatial foundation-model embeddings can support accurate binary cropland mapping while reducing dependence on task-specific modelling and large labelled datasets.

  • Method

    The study evaluates frozen annual AlphaEarth embeddings with lightweight classifiers, varying labelled-pixel amounts and training years, against CDL labels and an independent human consensus.

  • Results

    93.7% overall accuracy and 90.8% balanced accuracy were achieved on held-out Maine patches; 60,000 labelled pixels nearly matched 8.6 million, transfer held across 2018–2023, and the model agreed with human references at 95.3%.

  • Takeaways & Limitations

    Frozen geospatial embeddings are supported as a low-compute candidate for regional cropland mapping within the evaluated Maine setting.

  • Takeaways & Limitations

    Inference is limited by the single-state, forest-dominated setting, 30 m-derived reference labels aligned to a 10 m grid, spatial autocorrelation, same-region temporal testing, and one-block human validation.

Abstract

from arXiv · show

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.

1 Introduction

This paper examines whether frozen annual AlphaEarth embeddings support trustworthy, label-efficient binary cropland mapping with lightweight classifiers. It evaluates classifier performance, label needs, temporal transfer, embedding structure, and agreement with independent human interpretation in Maine.

  • 1 Introduction: Geospatial foundation models replace much manual feature engineering with reusable representations, but users may still need to fine-tune heavy backbones or attach task-specific heads.AlphaEarth and TESSERA instead release fixed-length per-pixel embeddings that can feed light classifiers without heavy model training or execution.
  • 1 Introduction: The task is binary separation of cultivated from non-cultivated land, with cultivated land operationally defined from crop classes rather than the USDA’s multi-year Cultivated Layer.The paper distinguishes crop, cropland, cultivated land, and non-cultivated land before defining the target.
  • 1 Introduction: Prior work reports promising embedding-based performance but also task dependence, limited spatial transferability, and weaker generalization across regions.These findings motivate testing temporal reuse and independent validation rather than relying solely on reference-product accuracy.
  • 1 Introduction: Existing evaluations largely report accuracy against a single reference product in one region and year, leaving embedding structure, cross-year reuse, reference correctness, and human agreement insufficiently examined together.The paper frames these omissions as the central gap motivating its integrated evaluation.
  • 1 Introduction: The study combines frozen AlphaEarth embeddings, lightweight classifiers, label-efficiency analysis, temporal transfer, embedding-space characterization, and independent human validation for cultivated-versus-non-cultivated mapping.The experiments use Maine, annual embeddings, and USDA Cropland Data Layer-derived labels while deliberately keeping the downstream classifier simple.
  • 1 Introduction: The paper explicitly limits label efficiency to CDL-labelled pixels, temporal transfer to the same Maine sampling frame, and human validation to one contiguous 2023 block.These definitions prevent pixel counts, cross-year results, and local human agreement from being interpreted as independent sites or geographic generalization.

2 Data and Study Area

The study uses annual AlphaEarth embeddings and binary labels derived from the USDA CDL across spatially separated, agriculture-enriched patches in Maine. Its data design preserves year-specific labels and reduces spatial leakage, while the 30 m reference aligned to a 10 m grid limits pixel-level interpretation.

  • 2.1 Maine as a study area: Maine is forest-dominated, with cultivated land covering only about 3% of the state in 2023 and concentrated in sparse, fragmented agricultural regions.This minority-class structure is a central difficulty and limits how directly the results may transfer to different agricultural landscapes.
  • 2.2 Ground truth: from 254 classes to a binary mask: Binary labels are built separately for 2018–2023 by grouping genuine crop classes from the annual 254-class Cropland layer into cultivated and assigning all other classes to non-cultivated.The annual Cropland layer is used instead of the multi-year Cultivated Layer because the latter does not represent an individual year.
  • 2.2 Ground truth: from 254 classes to a binary mask: Nearest-neighbour alignment transfers each 30 m CDL label to a 3 × 3 block of 10 m embedding pixels, forcing boundaries onto the coarser grid.Mismatch concentrates where 30 m cells straddle field edges, so CDL-based pixel accuracy is not direct evidence of 10 m boundary accuracy or independent 10 m reference information.
  • 2.3 Patch generation and sampling: About 22% of the enriched dataset is cultivated because patches are retained only when at least 10% of pixels are cultivated.This improves class balance for training and evaluation but does not create a probability-based sample representative of Maine statewide.
  • 2.3 Patch generation and sampling: 192 accepted patches are spatially separated by several kilometres, with interspersed training, validation, and test splits across Maine’s agricultural regions.The design prevents neighbouring-patch leakage while concentrating coverage in Aroostook County and central and southern valleys.
  • 2.4 The AlphaEarth embeddings: The dataset uses 64-dimensional annual AlphaEarth vectors as the sole per-pixel input for classification.Each vector summarizes a 10 m ground cell over a year and replaces raw reflectance features.
  • 2.4 The AlphaEarth embeddings: The three-channel embedding visualization preserves field boundaries and corresponds closely to cultivated areas in the binary label.The first three components explain about 57% of embedding variance, compared with about 5% for three arbitrary bands.
  • 2.3 Patch generation and sampling: The separate contiguous validation block is selected to maximize cultivated proportion rather than provide a random statewide sample.It therefore complements, but does not replace, the spatially separated patch design for assessing performance under continuous conditions.

3 Embedding-Based Classification Framework

The framework treats frozen 64-dimensional AlphaEarth embeddings as per-pixel representations read by lightweight classifiers, without fine-tuning the foundation model. It compares classifier complexity and training-pool size while evaluating spatially held-out patches.

  • 3.1 Problem set-up: The framework is designed around frozen embeddings because AlphaEarth is released as precomputed representations, making the embedding—not a newly trained deep model—the primary object of study.Per-pixel classification is justified by the embedding’s encoded spatial and temporal context, keeping the downstream model light.
  • 3.1 Problem set-up: Per-pixel 64-dimensional AlphaEarth embeddings are mapped to binary cultivated-versus-non-cultivated labels using lightweight classifiers, with spatially separated patches held out for testing.The embeddings are flattened into pixel-level rows, while the independent spatial unit for generalisation remains the held-out patch.
  • 3.2 Two training regimes: balanced sample against full pool: 8.6 million labelled training pixels are compared with a balanced 60,000-pixel sample, while the 20 test patches remain unused during training.The balanced regime uses 30,000 pixels from each class; the comparison measures downstream pixel-sample efficiency rather than independent manual annotations.
  • 3.3 Classifiers: The classifier ladder ranges from logistic regression through random forest and gradient boosting, with soft voting and stacking also tested under fixed, untuned settings.This design asks whether simple models perform nearly as well as more flexible ensembles and avoids attributing differences to unequal hyperparameter search.
  • 3.3 Classifiers: Logistic regression tests whether the embedding is close to linearly separable, whereas random forests and boosting represent nonlinear interactions and sequential error correction.The logistic model uses a flat boundary; random forests average independently fitted trees, while boosting fits successive trees to remaining errors.

4 Embedding-Space Analysis

The 64-dimensional AlphaEarth embeddings have non-Gaussian coordinate distributions but stable angular structure, with class signal distributed across the full vector. This geometry supports lightweight classification while making lossless dimension removal unsafe in this task-specific protocol.

  • 4.3 Normality and marginal shape: The embeddings are not well described by coordinate-wise Gaussian assumptions, with normality rejected across essentially all tested dimension-class combinations.126 of 128 combinations rejected normality on all three tests, and the normal distribution was never preferred by AIC.
  • 4.4 Per-dimension signal and class separability: 61 of 64 dimensions differed between classes in both location and spread, while 31 reached a large effect size and none was individually decisive.The class mean profiles overlap heavily, indicating that separability depends on distributed signal rather than one dominant coordinate.
  • 4.5 Angular geometry and a low-capacity supervised baseline: 90.2% accuracy from a nearest-class-centroid rule shows that embedding geometry alone captures substantial class separation before trained decision boundaries are added.The classes have within-class cohesion 0.808, cross-class similarity 0.599, a separation gap of 0.209, and centroid angle 42.2°.
  • 4.6 Can the 64 dimensions be compressed?: 92.3% accuracy was reached using all 64 principal components, up from 90.2% with two, even after variance had largely saturated.About 15, 22, and 43 components explain 90%, 95%, and 99% of variance, respectively; low-variance directions still carry class signal.
  • 4.6 Can the 64 dimensions be compressed?: Removing the three effectively negligible dimensions slightly reduced accuracy, from 92.96% to 92.89% for random forest and from 92.28% to 92.13% for logistic regression.The result is specific to this balanced Maine binary-classification protocol; smaller subsets may suffice when a small accuracy loss is acceptable.
  • 4.2 Directional geometry: Tangent-space projection changed accuracy by at most 0.03 percentage points, supporting the use of the original angular representation for subsequent classification.These analyses used a balanced 100,000-pixel sample with a 70/30 split and are not directly comparable to held-out-patch results.

5 Classification Results

Frozen AlphaEarth embeddings support accurate binary cropland mapping with little sensitivity to the lightweight classifier. Random forest provides 93.75% overall and 90.8% balanced accuracy, while class balancing improves minority recall at a modest overall-accuracy cost.

  • 5.1 The classifier ladder: 93.48% to 93.75% overall accuracy spanned the full-pool classifiers, only 0.27 percentage points, with macro-F1 around 0.89 throughout.Logistic regression was nearly as accurate as the random forest and XGBoost, indicating that classifier choice mattered little once embeddings were fixed.
  • 5.2 Class imbalance and per-class behaviour: About 82% overall accuracy would result from labeling every test pixel non-cultivated, so balanced accuracy is necessary to assess minority-class recovery.The model’s 90.8% balanced accuracy shows that cultivated pixels were recovered rather than sacrificed to the majority class.
  • 5.1 The classifier ladder: 92.4% to 93.0% balanced accuracy resulted from training on 60,000 balanced pixels, compared with 89.6% to 90.8% for full-pool training.Balancing lowered overall accuracy by about one percentage point while increasing cultivated-class recall from about 86% to about 93%.
  • 5.2 Class imbalance and per-class behaviour: 80.9% precision and 86.2% recall for cultivated pixels trailed the non-cultivated class, whose precision and recall both exceeded 95%.The chosen model’s cultivated F1 was 0.835 versus 0.961 for non-cultivated pixels.
  • 5 Classification Results: 93.75% overall accuracy and 90.8% balanced accuracy were achieved by the chosen random forest on held-out test patches.Non-cultivated recall was 95.4% and cultivated recall 86.2%, reflecting the harder, more heterogeneous minority class.
  • 5.3 Spatial predictions: The AlphaEarth predictions were nearly indistinguishable from ground truth on easier held-out patches, while remaining errors were spatially clustered.Figure 7 compares embedding false-colour imagery, CDL ground truth, random-forest predictions, and pixel-level agreement maps.

6 Label Efficiency

A balanced sample of 60,000 labelled pixels captures nearly all available signal, while much larger full-pool training adds little and can reduce imbalance-aware performance. Classifier combinations likewise provide only marginal gains over strong lightweight models.

  • Sample-size comparison: 93.75% overall accuracy from the full ∼8.6 million-pixel pool exceeds the balanced 60,000-pixel sample by only about 1.3 percentage points.Other models show similar behaviour, indicating diminishing returns from the much larger training pool.
  • Sample-size comparison: 93.0% balanced accuracy from the balanced-sample support vector machine and 92.6% from the balanced random forest exceed the full-pool random forest’s 90.8%.The natural 78:22 class ratio in the full pool pulls the decision boundary toward the majority non-cultivated class.
  • Ensemble comparison: 92.64% from the stacked ensemble is only 0.19 percentage points above the balanced random forest and below the full-pool random forest.Learning to combine four classifiers contributes less than adding data, while the embedding already supplies most useful discriminative information.
  • Representation efficiency: 90.2% from a nearest-centroid rule using only two labelled class centroids supports the conclusion that the task is not primarily limited by labelled-pixel volume.The result is consistent with most discriminative information already being present in the embedding representation.
  • Scope: A balanced 60,000-pixel sample should be read as pixel-sample efficiency, not a minimum requirement or 60,000 independent annotation sites.The pixels are spatially correlated, and the conclusion is limited to agriculture-enriched held-out patches and the default threshold.

7 Cross-Year Transfer

Annual AlphaEarth embeddings support same-region temporal transfer: models trained in one year remain accurate across Maine data from 2018 to 2023. The result indicates regional temporal stability, with 2019 a modest unexplained low point rather than a demonstrated cloud-cover effect.

  • Transfer performance: 92.5%–93.8% overall accuracy across 2018–2023 shows little loss when a random forest trained in one year is applied unchanged to another.Off-diagonal transfer entries remain close to matched diagonal performance across the full train-year by evaluation-year matrix.
  • Transfer performance: 91.8%, 90.6%, 91.5%, 90.8%, 90.8% and 90.8% are the matched balanced accuracies for 2018–2023, respectively.Across all train-evaluate combinations, balanced accuracy remains approximately 86%–92%, indicating consistent minority-class recovery.
  • Year-specific variation: 2019 is the weakest year both on the matched diagonal and as a transfer target, with the 2023 model reaching about 92.5% there.The authors do not attribute this small dip to cloud cover because full-year clear-sky availability does not align with the accuracy pattern.
  • Scope: The experiment establishes same-region temporal transfer within reused Maine patch locations, not geographic transfer or general temporal invariance.Application to new regions still requires independent validation, and annual retraining may not be necessary for this regional workflow.

8 Independent Human Validation

Independent photo-interpretation on 385 points in one contiguous 2023 Maine block shows the AlphaEarth-plus-random-forest map agrees with human consensus more closely than the CDL. Its advantage over the CDL is significant, while its difference from fine-tuned TerraMind is not.

  • Validation design: The human reference was built independently by two interpreters blind to CDL labels, using a simple random sample of 385 valid pixels in one spatially separate 2023 block.The interpreters agreed on 91.2% of points before reconciliation, with κ=0.63.
  • Model versus human reference: 95.3% agreement with human consensus for AlphaEarth plus random forest exceeds 91.7% for the CDL, with κ=0.82 versus 0.72.The model’s cultivated-class user’s accuracy is 86%, compared with 69% for the CDL, consistent with fewer cultivated false positives.
  • Paired significance testing: p=0.016 for the exact paired test shows the model agrees significantly more closely with human consensus than the CDL on the 385 interpreted points.The model is correct on 22 discordant points where the CDL is wrong, versus 8 in the reverse direction.
  • Comparison with TerraMind: 93.5% overall agreement for fine-tuned TerraMind is below the model’s 95.3%, but the difference is not significant (p=0.14).TerraMind is above the CDL overall, though its balanced accuracy is 87.5%, just below the CDL’s 89.1%.
  • Scope: The result is indicative rather than a definitive ranking because it uses one block and the same two interpreters, despite showing the lightweight approach is not obviously behind TerraMind.The comparison does not establish statewide CDL correction or a controlled computational-cost comparison.

9 Discussion

The discussion frames frozen-embedding mapping as a low-compute regional prototype whose apparent advantage over CDL labels is promising but locally bounded. Operational deployment still requires broader validation, calibration, uncertainty assessment, and monitoring.

  • Why a classifier can be closer to the truth than its own training labels: The classifier can smooth some isolated CDL errors because embedding-consistent regularities are learned, whereas individual reference misassignments are not regularities.This mechanism applies only when CDL noise is not systematically aligned with the embedding.
  • Operational implications: The human-reference result is limited to 385 points in one contiguous 2023 block and cannot characterize CDL error across Maine or across years.The paired test reaches significance locally, but the sample does not establish statewide correction of the reference product.
  • Operational implications: The workflow avoids foundation-model training and uses a standard downstream classifier, but statewide use would require probability-based validation, prevalence-aware calibration, area-adjusted accuracy, uncertainty estimates, and shift monitoring.The paper presents this as a reproducible regional baseline rather than a deployment-ready statewide product.
  • Operational implications: Generalization is constrained by Maine’s forest-dominated, fragmented landscape, a nonrepresentative agriculture-enriched sample, and a 30 m reference aligned to a 10 m embedding grid.Spatial autocorrelation also reduces the effective number of independent observations, although separated test patches limit direct train-test leakage.
  • Operational implications: Frozen embeddings are not shown to outperform fine-tuning: TerraMind differs nonsignificantly, although the frozen route shifts the practical comparison toward cost and public availability.AlphaEarth is released as embeddings only, so users cannot fine-tune its closed model.

10 Conclusions

The conclusions find that frozen AlphaEarth embeddings support accurate, label-efficient binary cropland mapping with lightweight classifiers. Temporal reuse within Maine and closer agreement with local human interpretation support a competitive regional workflow, while broader comparisons and spatial context remain future needs.

  • 10 Conclusions: 93.7% overall accuracy and 90.8% balanced accuracy were achieved on held-out Maine patches using frozen 64-dimensional AlphaEarth vectors and a lightweight classifier.The task is binary cultivated-versus-non-cultivated classification using USDA CDL-derived labels.
  • 10 Conclusions: Logistic regression reached 93.48% versus 93.75% for gradient boosting, and nearest-class-centroid classification reached 90.2%.These results indicate that classifier choice matters relatively little once the embedding is fixed.
  • 10 Conclusions: A balanced 60,000-pixel sample captured most discrimination from roughly 8.6 million pixels, while embedding signal remained distributed across many dimensions.The embedding analysis found unit-hypersphere vectors, non-normal coordinates, and no dimension that could be discarded without measurable loss.
  • 10 Conclusions: 95.3% agreement with human consensus exceeded CDL agreement at 91.7%, while remaining statistically indistinguishable from TerraMind at 93.5%.The model-versus-CDL difference was significant at p = 0.0161; the TerraMind comparison had p = 0.14.
  • 10 Conclusions: A classifier fitted in 2023 transferred across 2018–2023 within the same Maine sampling frame with little loss.This conclusion concerns temporal reuse within the same region, not geographic transfer.
  • 10 Conclusions: Future work should compare openly released embeddings directly, broaden fine-tuned segmentation tests, and add light spatial context for boundary and small-parcel errors.The proposed comparisons should use contiguous tiles and retain paired testing where appropriate.

Data and Code Availability

The analysis pipeline and source-data references are openly documented, with code, notebooks, and reproducibility wrappers available under the MIT licence.

  • Data and Code Availability: The full pipeline is available under the MIT licence in the alphaearth-cropland-maine GitHub repository.The repository includes the ae Python package, figure and table notebooks, and wrappers for classification and cross-year runs.
  • Data and Code Availability: AlphaEarth embeddings and CDL-derived labels come from specified Google Earth Engine collections, while patch extraction and photo-interpretation use named open tools.The 385-point consensus reference is available from the corresponding author on reasonable request.

Funding

The study was partially funded through the Space It Up project by the Italian Space Agency and the Ministry of University and Research.

  • Funding: The Space It Up project provided partial funding through ASI and MUR under contract 2024-5-E.0.The funding record also lists CUP n. I53D24000060005.

Disclaimer

The paper states that its views belong to the authors and may not reflect the views or policies of the Food and Agriculture Organization of the United Nations.

  • The authors disclaim representing the views or policies of the Food and Agriculture Organization of the United Nations.
Loading 2609.17138v1…