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

Mohammadreza Narimani, Alireza Pourreza, and Parastoo Farajpoor

ASABE Annual International Meeting 2026 · Indianapolis, Indiana · Poster presented July 14, 2026

Digital Agriculture Laboratory · Department of Biological and Agricultural Engineering · University of California, Davis

Presented at ASABE 2026

I presented this research at the ASABE Annual International Meeting 2026 in Indianapolis, Indiana. The poster brought together geospatial foundation models, deep learning, and uncertainty-aware mapping to explore a practical question: can analysis-ready satellite embeddings identify processing tomato fields accurately at statewide scale?

Mohammadreza Narimani standing beside his AlphaEarth tomato mapping poster at ASABE 2026 in Indianapolis
Mohammadreza Narimani with the ASABE 2026 poster in Indianapolis.
Mohammadreza Narimani explaining AlphaEarth geospatial embeddings and tomato crop mapping to an attendee at ASABE 2026
Discussing the AlphaEarth and U-Net workflow during the poster session.

Why Statewide Tomato Mapping Matters

Reliable field-scale crop maps support acreage estimation, water accounting, yield forecasting, harvest logistics, contract planning, and agricultural policy. Many statewide products are retrospective, while conventional remote-sensing pipelines often depend on cloud masking, atmospheric correction, gap filling, hand-engineered vegetation indices, and case-specific temporal features.

We tested whether Google DeepMind's AlphaEarth Foundations could provide a more direct starting point. AlphaEarth represents every 10 m pixel with a 64-dimensional annual embedding learned from multi-source Earth observation data. Instead of interpreting individual bands as physical measurements, the downstream model uses the complete embedding as an analysis-ready description of land-surface behavior.

Research objective: map processing tomato versus non-tomato fields from annual 64-band AlphaEarth chips using a U-Net, then quantify where predictions are reliable and where uncertainty is concentrated.

Data and Spatially Independent Evaluation

LandIQ 2018 crop polygons provided the reference labels. We assembled a balanced dataset of 4,742 processing tomato fields and 4,742 non-tomato fields. The negative class included alfalfa, wheat, corn, beans, orchards, and other crops so the model had to learn tomato-specific patterns rather than a simple tomato-versus-one-crop contrast.

Fields were partitioned spatially before training to reduce leakage from nearby locations. The final split included 6,638 training fields, 1,422 validation fields, and 1,424 test fields. Invalid and NoData pixels were masked during both optimization and evaluation.

9,484Total field polygons
64AlphaEarth embedding bands
10 mSpatial resolution
6,638Training fields
1,422Validation fields
1,424Independent test fields

End-to-End Workflow

The pipeline moves from statewide crop polygons to balanced field sampling, 64-band embedding extraction, spatial splitting, cloud-based U-Net training, pixel-wise segmentation, uncertainty estimation, and field-level aggregation. Model development ran on AWS SageMaker using an NVIDIA A10G GPU.

End-to-end AlphaEarth tomato mapping workflow from LandIQ crop polygons through U-Net training, uncertainty estimation, and farm-level aggregation
Figure 1. End-to-end workflow for dataset construction, segmentation, uncertainty estimation, and farm-level summaries.

A 64-Channel U-Net with Uncertainty Estimation

The fully symmetric U-Net accepts the complete 64 × H × W AlphaEarth tensor and produces an aligned pixel-wise probability map. The loss combines masked binary cross-entropy with soft Dice loss, balancing class discrimination with boundary overlap.

To move beyond a single hard prediction, dropout remained active during inference. Each field chip passed through the network 100 times; the predictive mean became the final soft map, while pixel-wise variance represented model uncertainty. This makes field edges, mixed pixels, and ambiguous geometry visible for review.

Fully symmetric 64-channel U-Net architecture with AlphaEarth input tensor, segmentation mask, probability map, and Monte Carlo dropout uncertainty map
Figure 2. The 64-channel U-Net produces aligned segmentation, probability, and Monte Carlo dropout uncertainty maps.

Results on Unseen Locations

The model retained high performance on the spatially independent test set. It accurately separated tomato fields from a diverse set of agricultural classes while tracing field boundaries with limited confusion.

Test metricScore
Pixel accuracy99.19%
Precision98.69%
Recall99.40%
F1 score99.04%
Intersection over Union98.11%
Chip accuracy99.02%
Representative tomato and non-tomato AlphaEarth chips with binary masks, mean predicted probabilities, and Monte Carlo dropout variance
Figure 3. Representative tomato and non-tomato chips. Uncertainty is low in field interiors and increases near boundaries and irregular features.

What the Results Mean

AlphaEarth embeddings preserve crop-relevant spatial and temporal structure in a compact representation. This lets the segmentation network focus on tomato-specific field patterns and boundaries instead of first reconstructing a large preprocessing pipeline from raw observations.

Practical takeaway: geospatial foundation-model embeddings can support accurate field-scale crop mapping without manual spectral feature engineering, while uncertainty maps identify locations where human review or auxiliary data may still be valuable.

The current study uses labels from one year and a balanced evaluation dataset. Future work will test multi-year labels, more crop classes, realistic statewide class imbalance, transfer beyond California, and additional tasks such as crop stress, disease, and management-zone mapping.

Explore the ASABE 2026 Poster

The full poster provides a visual summary of AlphaEarth foundations, the spatial split, embedding-space analysis, U-Net architecture, discriminant embedding bands, prediction examples, uncertainty maps, and the main findings.

Mohammadreza Narimani ASABE 2026 research poster about mapping California processing tomato fields with AlphaEarth geospatial embeddings, U-Net deep learning, and uncertainty estimation
ASABE 2026 poster: AlphaEarth geospatial embeddings and deep learning for field-scale processing tomato mapping in California.

Paper and Citation

Narimani, M., Pourreza, A., & Farajpoor, P. (2026). Mapping Tomato Cropping Systems in California Using AlphaEarth Geospatial Embeddings and Deep Learning Analysis. arXiv preprint arXiv:2605.21804.

Contact

Mohammadreza Narimani
PhD Candidate, University of California, Davis

📧 mnarimani@ucdavis.edu  |  🎓 Google Scholar