betaearth-segformer-film-robust

BetaEarth SegFormer-B2 with FiLM conditioning — robust variant trained with curriculum modality dropout for single-modality deployment.

Part of the BetaEarth family. Unlike the frozen+FiLM models which perform best with 4 input modalities, this model handles any subset of {S2 L1C, S2 L2A, S1 RTC, COP-DEM} gracefully.

Results

All-modality (full input):

Metric	Value
Val cosine similarity (all)	0.878
Total parameters	104.8M
Trainable parameters	104.8M
Inputs	S2 L1C+L2A (9ch), S1 RTC (2ch), COP-DEM (1ch), DOY
Output	(H, W, 64) float32, L2-normalised

Single-modality robustness (main contribution):

Input subset	Cosine sim	Notes
All modalities	0.878	Comparable to frozen+FiLM (0.886)
L1C + DEM	0.850
L2A + DEM	0.823
S2 both (L1C+L2A)	0.818
L1C only	0.806	Deployable with just L1C
L2A only	0.755	Up from 0.537 in frozen+FiLM
S1 only	0.712	Up from ~0.60 in frozen+FiLM
DEM only	0.609

Training

• Base: reinit_fusion checkpoint (0.886 test cos sim)
• Strategy: curriculum dropout — 60% single modality (15% each), 20% random pair, 20% all
• Duration: 20 epochs, AdamW, cosine LR
• W&B: r4ctfspm

Usage

pip install betaearth

from betaearth import BetaEarth

model = BetaEarth.from_pretrained("asterisk-labs/betaearth-segformer-film-robust")

# Full input — all modalities
embedding = model.predict(
    s2_l1c=s2_l1c,   # (9, H, W) uint16
    s2_l2a=s2_l2a,   # (9, H, W) uint16
    s1=s1,            # (2, H, W) float32
    dem=dem,          # (1, H, W) float32
    doy=182,
)

# Single modality — works without all inputs
embedding = model.predict(s2_l2a=s2_l2a, doy=182)

All BetaEarth models

Model	Cos Sim	Params	Best for
betaearth-segformer-film	0.886	0.3M	Best all-modality quality
betaearth-segformer-film-robust	0.878	104.8M	Flexible deployment
betaearth-segformer-film-hilr	0.886	0.3M	Alt frozen
betaearth-segformer	0.880	104.8M	No timestamp
betaearth-segformer-film-scratch	0.883	104.8M	End-to-end
betaearth-rgb-only	0.836	26.3M	Minimal data

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

CC-BY 4.0. Training data attribution: "The AlphaEarth Foundations Satellite Embedding dataset is produced by Google and Google DeepMind."