betaearth 0.2.0

BetaEarth: AlphaEarth Embedding Emulator

Embedding Sentinel-2 and Sentinel-1 with a Little Help of AlphaEarth

---

What is BetaEarth?

BetaEarth is an open-source model that produces dense 10m geospatial embedding fields from Sentinel-2 and Sentinel-1 imagery. It is trained to reproduce the outputs of AlphaEarth Foundations (AEF) — a closed-source embedding model released by Google and Google DeepMind — using only AEF's publicly available precomputed embeddings as supervision.

BetaEarth has no access to AEF's weights or architecture. It is an independent model, not a variant or extension of AEF.

Why does this matter?

• Reproducibility: AEF embeddings cannot be generated for new data without Google Earth Engine access. BetaEarth can run locally on any Sentinel-2/S1 imagery.
• Auditability: BetaEarth enables the community to probe a closed-source model's behaviour — identifying biases, modality sensitivities, and failure modes — without direct model access.
• Security research: This work demonstrates that releasing embeddings may not be a risk-free alternative to releasing model weights.

---

Models

We release 8 model variants spanning different trade-offs between quality, parameter efficiency, and input requirements.

Main results (6,200-tile test set)

Model	Test Cos Sim	Std	LULC Acc	Model Size	Inputs
SF curriculum (robust)	0.873	0.109	0.833	104.8M	Any subset of S2/S1/DEM + DOY
SF frozen+FiLM (reinit)	0.886	0.098	0.873	104.8M	S2 L1C+L2A, S1, DEM, DOY
SF frozen+FiLM (hilr)	0.886	0.099	0.866	104.8M	S2 L1C+L2A, S1, DEM, DOY
SF from scratch+FiLM	0.883	---	0.835	104.8M	S2 L1C+L2A, S1, DEM, DOY
SF no FiLM (ISPRS)	0.880	0.101	0.869	104.8M	S2 L1C+L2A, S1, DEM
DINOv3 ViT-L/16 (sat)	0.874	0.100	0.870	304M	6 primitives + DOY
DINOv3 ViT-S/16 (nat)	0.861	0.109	0.863	23.8M	6 primitives + DOY
SF RGB-only+FiLM	0.836	---	0.823	26.3M	S2 RGB, DOY
Real AlphaEarth (ceiling)	---	---	0.889	---	---

The curriculum (robust) model handles any modality subset gracefully:

Input subset	Cosine sim
All modalities	0.873
L1C only	0.806
L2A only	0.755
S1 only	0.712
DEM only	0.609

Which model should I use?

Use case	Recommended model	Why
General use (default)	SF curriculum (robust)	Works with any input subset; best for real-world deployment
Maximum quality	SF frozen+FiLM (reinit)	Highest cos sim (0.886) — requires all 4 modalities
No timestamp needed	SF no FiLM (ISPRS)	Does not require day-of-year input; still achieves 0.880
Lightweight / edge	DINOv3 ViT-S/16	23.8M params, good quality (0.861)
Minimal data requirements	SF RGB-only+FiLM	Only needs 3-band RGB + day-of-year
Research / ablation	SF frozen+FiLM (hilr)	Alternative fusion strategy for comparison

Architecture overview

DINOv3 models use a single shared frozen DINOv3 backbone applied to 3-band spectral primitives:

Primitive	Bands	Captures
True-colour RGB	B04/B03/B02	Visual texture, built environment
False-colour IR	B08/B04/B03	Vegetation health (NIR)
SWIR composite	B12/B11/B04	Moisture, bare soil, burn scars
Red-edge	B07/B06/B05	Canopy structure, chlorophyll
SAR	VV/VH/ratio	Structure, moisture (from S1)
Topography	Elevation/Slope/Aspect	Terrain (from COP-DEM)

Primitives are fused via permutation-invariant cross-attention (SetFusion).

SegFormer models use 4 separate MiT-B2 encoders processing each modality's raw bands natively (9ch S2-L1C, 9ch S2-L2A, 2ch S1, 1ch DEM), with channel concatenation fusion.

All models use FiLM temporal conditioning (day-of-year modulation) except the ISPRS baseline.

Key findings

• Temporal conditioning as spectral compensation: FiLM importance scales inversely with spectral access — RGB-only (22pp) > DINOv3 (18pp) > SegFormer scratch (14pp) > frozen SegFormer (5pp).
• Multi-temporal averaging of 4+ observations improves emulation by up to +18pp over single timestamps.
• Predicted embeddings retain 98% of downstream LULC classification accuracy and are robust to 32x compression.

---

Model Properties

Property	Value
Output	Dense embedding field — (H, W, 64) per tile at 10m resolution
Output normalisation	L2-normalised per pixel (unit vectors on S^63)
Quantisation	Original AEF: int8 on S^63; BetaEarth outputs float32
Tile size	10.68 x 10.68 km (1068 x 1068 px), Major TOM grid
Training data	62,489 Major TOM grid cells (49,991 train / 6,248 val / 6,250 test)
Loss	Cosine similarity + 0.1 * MSE, masked to valid pixels

---

Quickstart

pip install betaearth

from betaearth import BetaEarth

model = BetaEarth.from_pretrained()  # default: robust variant
# BetaEarth(params=104.8M, device=cuda)

# All inputs are raw (unnormalised) — preprocessing is handled internally
embedding = model.predict(
    s2_l2a=s2_l2a,   # (9, H, W) uint16 DN (~0-10000)
    s2_l1c=s2_l1c,   # (9, H, W) uint16 DN (~0-10000)
    s1=s1,            # (2, H, W) float32 linear power
    dem=dem,          # (1, H, W) float32 elevation in meters
    doy=182,          # day of year (1-366)
)
# embedding: (H, W, 64) float32 numpy array, L2-normalised per pixel

Any modality can be omitted — the model handles missing inputs via zeroed features:

# S2-only (no S1, no DEM)
emb = model.predict(s2_l2a=s2_l2a, doy=182)

# S2 + DEM, no S1
emb = model.predict(s2_l2a=s2_l2a, dem=dem, doy=182)

Multi-temporal averaging

import numpy as np

preds = []
for s2, s1, doy in zip(s2_timeseries, s1_timeseries, doys):
    pred = model.predict(s2_l2a=s2, s1=s1, dem=dem, doy=doy)
    preds.append(pred)

# Simple averaging — saturates at ~4 observations
annual = np.mean(preds, axis=0)
annual /= np.linalg.norm(annual, axis=-1, keepdims=True)

---

Data Access

All training data is from the Major TOM community project and is freely available on HuggingFace:

Dataset	Description
Major-TOM/Core-S2-L2A	Sentinel-2 L2A imagery
Major-TOM/Core-S2-L1C	Sentinel-2 L1C imagery
Major-TOM/Core-S1-RTC	Sentinel-1 RTC imagery
Major-TOM/Core-AlphaEarth-Embeddings	AEF target embeddings

Data normalisation

All input data should be stored as raw values. Normalisation happens inside the model:

• S2 L1C/L2A: uint16 DN (0-10000+), divided by 10000 internally
• S1 RTC: linear power (float32, ~0-200), log-transformed internally
• COP-DEM: pre-normalised to [0, 1] before passing to the model

Important: S2 band order must follow Major TOM convention: [B02, B03, B04, B08, B05, B06, B07, B11, B12] (10m bands first, then 20m).

---

Reproduce

git clone https://github.com/asterisk-labs/betaearth
cd betaearth
conda env create -f environment.yml
conda activate betaearth

# Train (requires A100 GPU)
python train_multi.py --batch_size 8 --max_epochs 20

# SegFormer FiLM variants (frozen encoder)
python train_segformer_frozen_variants.py --ckpt checkpoints/isprs_v1_epoch19_0.875.ckpt

# Evaluate on test set
python run_evaluation.py --ckpt checkpoints/multi_final.ckpt

# Generate paper figures
python generate_figures.py
python plot_training_curves.py

# Multi-temporal experiments
python test_multitemporal.py --ckpt checkpoints/segformer_film_frozen_best.ckpt

See CHECKLIST.md for a full step-by-step guide to reproducing each experiment in the paper.

---

Citation

@inproceedings{czerkawski2026betaearth,
  title     = {BetaEarth: Emulating Closed-Source Earth Observation Foundation Models Through Their Public Embeddings},
  author    = {Czerkawski, Mikolaj},
  booktitle = {ISPRS Congress 2026},
  year      = {2026}
}

---

License and Attribution

BetaEarth model weights are released under CC-BY 4.0, matching the license of the AlphaEarth Foundations embedding archive used for training supervision.

Required attribution for AEF training data:

"The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind."

Training imagery is sourced from Major TOM (Apache 2.0) and Copernicus Sentinel (free and open access).