Mughees et al. evaluate AlphaEarth embeddings for cropland mapping in Maine, USA; 93.7% overall accuracy on held-out patches
From Foundation Embeddings to Cropland Maps: Label Efficiency, Temporal Transferability and Independent Human Validation
From Foundation Embeddings to Cropland Maps: Label Efficiency, Temporal Transferability and Independent Human Validation Mohammad Ammar Mughees1, Giovanni Montefoschi1, Zhongxin Chen2, and Maria Antonia Brovelli1 1Department of Civil and Environmental Engineering, Politecnico di Milano, Milan, Italy 2Food and Agriculture Organization of the United Nations, Rome, Italy Abstract. Geospatial foundation models provide reusable representations of satellite imagery that can 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 image patches and labels derived from the USDA Cropland Data Layer (CDL). Without fine-tuning the foundation model, a lightweight classifier reaches approximately 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 labelled class centroids but no iterative parameter fitting, reaches 90.2%. A balanced sample of 60,000 labelled pixels is within approximately 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 temporal-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. Keywords: geospatial foundation models; AlphaEarth embeddings; cropland classification; Cropland Data Layer; label efficiency; random forest; photo-interpretation validation 1 Introduction Producing land-use and land-cover information used to be slow and expensive. It relied on the manual inter- pretation of imagery and on task-specific models that had to be built almost from scratch for every new prob- lem, and the biggest obstacle was data: training such models needed very large, carefully labelled datasets, and high-quality labels are scarce because they are costly to collect [7]. Satellite missions such as Sentinel and Landsat have been both an asset and a challenge in this respect. They provide a continuous stream of observations from which a dataset can be built, but the volume of that data is too large to use directly, so a geospatial expert has to spend considerable effort cleaning, compositing and engineering features before the data becomes usable [11]. Compositing, the usual way of dealing with clouds, also discards part of the temporal phenological signal that matters most for crops. Geospatial foundation models have begun to change this picture. Instead of engineering features by hand, these models learn the features themselves, mostly through self-supervised pre-training, and a single pre- trained model can then serve several downstream tasks. They divide into two broad ways of working. One route provides a pre-trained backbone that the user still has to fine-tune for each task, as with Prithvi [13] or Ter- raMind [14]. The other route runs the model once, at the developer’s side, and releases its output directly as embeddings: fixed-length per-pixel vectors that behave, in geospatial terms, like bands. AlphaEarth [7] and TESSERA [11] follow this second route, releasing 64- dimensional and 128-dimensional annual embeddings respectively. For every 10 m × 10 m pixel the user ob- tains one such vector, which compactly summarises a whole year of multimodal observations at that location, and which can be fed straight into a light classifier with no heavy model to train or run. This paper is concerned with that second family, and specifically with what it delivers for cropland mapping. Four related terms are used here in a specific sense. A crop is a cultivated plant grown for food, fibre or fodder. Cropland is land used to grow such crops, comprising arable land together with land under permanent crops [12]. Cultivated land is the broader category of land worked for agriculture; in the USDA Cropland Data Layer it is represented by a dedicated Cultivated Layer, 1 arXiv:2609.17138v1 [cs.CV] 15 Sep 2026
...