Modelpublic

asterisk-labs/betaearth-dinov3-vitl16

sourceHugging Facecc-by-4.0updated 6mo agoView on Hugging Face

0likes21downloads

Model Card

betaearth-dinov3-vitl16

BetaEarth DINOv3 ViT-L/16 (satellite pretrained, frozen) — best LULC downstream

Metric	Value
Test cosine similarity	0.874
LULC downstream accuracy	0.87
Trainable parameters	2.1M
Total parameters	304M (backbone loaded from torch.hub)
Inputs	6 spectral primitives + DOY

Note: This repo contains only the trainable weights (DPT head, FiLM, fusion, decoder — 2.1M). The frozen DINOv3 backbone is loaded from via at runtime.

Usage

All BetaEarth models

Model	Cos Sim	Trainable	Architecture
betaearth-segformer-film	0.886	0.3M	SegFormer frozen+FiLM
betaearth-dinov3-vitl16	0.874	2.1M	DINOv3 ViT-L/16
betaearth-dinov3-vits16	0.861	1.8M	DINOv3 ViT-S/16
betaearth-segformer	0.880	104.8M	SegFormer no FiLM
betaearth-rgb-only	0.836	26.3M	SegFormer RGB-only

License

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