From Foundation Embeddings to Cropland Maps: Label Efficiency, Temporal Transferability and Independent Human Validation
Authors: Mohammad Ammar Mughees, Giovanni Montefoschi, Zhongxin Chen, Maria Antonia Brovelli
Abstract
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.
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.
Geospatial foundation models pretrained on satellite imagery promise broad generalization across remote sensing tasks and regions, but their geographic transferability has not been systematically tested, especially in agriculture applications. This paper presents a controlled benchmark that evaluates three models, Prithvi, SpectralGPT, and SatMAE, on multi-temporal crop segmentation and change detection across four U.S. states, Iowa, North Carolina, California, and Minnesota. By assigning each train, validation, and test split to a separate region, we measure how well each model transfers to land it has not seen. All three degrade sharply under regional distribution shift, predicting only the most common crops while missing rare ones. We further find that fitting these models to a shared input format affects each one differently, which complicates direct architectural comparison. These results expose key limitations of current geospatial foundation models for agriculture and point to region aware evaluation as a necessary standard.
Machine learning, and deep networks in particular, are increasingly used to derive higher-level Earth observation (EO) products such as annual land-cover and crop-type maps. Many are generated operationally: each year a new acquisition is processed, typically with the same model, extending a multi-year archive. In the process these systems accumulate two kinds of useful signal that are almost never fed back into the model: the system's own archive of past predictions, and ancillary layers produced by other partners in a processing consortium. Both are normally used outside the network, as rule-based post-processing or a fixed input mask. Using the Copernicus Land Monitoring Service High Resolution Layer (HRL) Croplands crop-type product as a testbed, we show that bringing both signals inside the model turns a single-year, single-task pixel classifier into one that reasons across years. We introduce a Crop Type (CTY) embedding encoder that represents each past prediction as a confidence-scaled, time-ordered categorical token and attends over the year axis, and we study how the externally provided Base Vegetation Layer (BVL) mask should be represented in the model's inputs and outputs. To compare designs fairly when they relabel non-crop pixels, we evaluate on the 18 crop classes only and report precision and recall separately. On a pan-European dataset of about 5.4M labelled pixels, adding the prediction history raises crop-only F1 by 1.6 percentage points (pp) and, more importantly, corrects a recall-skewed error profile, with the largest gains on perennial and tree crops (olives +4.6, fruits +3.7, nuts +3.2 pp). Representing the BVL mask consistently in both the history and the target year adds about 2.5 pp on the crop classes. The approach is a low-cost recipe for any recurring geospatial or foundation model that emits class maps.
Syed Roshaan Ali Shah, Kasper Bonte, David Bekaert +2