Google DeepMind's AlphaEarth achieves 99% accuracy mapping California tomato fields

A team of researchers from the University of California and other institutions have demonstrated that Google DeepMind's AlphaEarth geospatial embeddings can replace hand-engineered spectral features for field-scale crop mapping. Using a balanced dataset of 4,742 tomato and 4,742 non-tomato fields from California's LandIQ 2018 crop polygons, they extracted 64-band AlphaEarth embedding chips and trained a U-Net segmentation model on AWS SageMaker. The model achieved 99.19% pixel accuracy, 98.69% precision, 99.40% recall, and 99.04% F1 score on a spatially independent test set of 1,424 fields. Uncertainty maps, generated via 100 Monte Carlo dropout inferences per chip, showed highest uncertainty near field edges and low uncertainty within interior regions. The results confirm that AlphaEarth embeddings retain crop-relevant spatial and temporal structure, enabling accurate crop mapping without manual feature engineering or repeated preprocessing across years.

The study, submitted as a preprint to ASABE 2026 AIM, represents a significant step toward analysis-ready geospatial pipelines for agricultural monitoring. Unlike traditional remote-sensing workflows that require handcrafted spectral indices and frequent recalibration, AlphaEarth embeddings provide a generalizable foundation that can be fine-tuned for specific crops and regions. The high intersection-over-union (98.11%) and chip-level accuracy (99.02%) suggest the approach is robust enough for operational use in supply-chain forecasting and policy planning. The researchers note that the method could be extended to other crop types and geographic areas, potentially reducing the cost and complexity of statewide crop mapping initiatives.