TESSERA v2: Scaling Pixel-wise Earth Foundation Models

Hugging Face Daily Papers Papers
earth-observation foundation-models scaling-study pixel-wise model-distillation self-supervised-learning computer-vision

Summary

The paper presents the largest controlled scaling study for Earth-observation foundation models, showing that pretraining loss poorly predicts downstream performance and providing an optimal compute allocation rule. It trains scaled pixel-wise models (0.5B and 1B parameters) and distills them into compact student models that outperform larger open and proprietary models.

Pixel-wise Earth-observation (EO) foundation models are now achieving state-of-the-art performance via generated spatial embeddings. However, how these models scale and how best to spend a pretraining budget remain poorly understood. We present the largest controlled scaling study for EO to date: 395 training runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each evaluated on 15 downstream tasks. We find that pretraining loss barely predicts downstream performance (|Pearson r| < 0.2), so selecting models by loss wastes a large share of the compute. We also find that, as the training budget grows, the encoder and the data should grow together while the projector stays fixed, which gives a simple rule for allocating compute. Using this rule, we train a family of pixel-wise models (0.5B and 1B, with a 2B model in training) and distill them into compact students for embeddings-as-data deployment. The 21-million-parameter distilled TESSERA v2-1B-M in aggregate outperforms all open and proprietary models tested, some of which are orders of magnitude larger. These students produce Matryoshka representations that are inexpensive to serve: a 16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 of the storage. Upon completion of training we plan to release v2 global embeddings covering 2017-2025. Together, these results give a concrete, empirically grounded recipe for scaling pixel-wise EO foundation models: train large encoders, select by downstream performance, and distil into flexible student models. All code will be released at https://github.com/ucam-eo/tessera.
Original Article
View Cached Full Text

Cached at: 07/09/26, 11:41 PM

Paper page - TESSERA v2: Scaling Pixel-wise Earth Foundation Models

Source: https://huggingface.co/papers/2607.03949 Authors:

,

,

,

,

,

,

,

,

,

,

,

Abstract

Large-scale controlled experiments reveal optimal scaling strategies for Earth-observation foundation models, enabling efficient training and deployment through encoder growth, downstream performance selection, and model distillation.

Pixel-wise Earth-observation(EO)foundation modelsare now achieving state-of-the-art performance via generatedspatial embeddings. However, how these models scale and how best to spend apretraining budgetremain poorly understood. We present the largest controlledscaling studyfor EO to date: 395 training runs on 1,024 GH200 superchips within a fixed pixel-wiseBarlow Twinsfamily, each evaluated on 15downstream tasks. We find that pretraining loss barely predicts downstream performance (|Pearson r| < 0.2), so selecting models by loss wastes a large share of the compute. We also find that, as the training budget grows, theencoderand the data should grow together while theprojectorstays fixed, which gives a simple rule for allocating compute. Using this rule, we train a family of pixel-wise models (0.5B and 1B, with a 2B model in training) and distill them into compact students for embeddings-as-data deployment. The 21-million-parameter distilled TESSERA v2-1B-M in aggregate outperforms all open and proprietary models tested, some of which are orders of magnitude larger. These students produceMatryoshka representationsthat are inexpensive to serve: a 16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 of the storage. Upon completion of training we plan to release v2 global embeddings covering 2017-2025. Together, these results give a concrete, empirically grounded recipe for scaling pixel-wise EOfoundation models: train largeencoders, select by downstream performance, and distil into flexible student models. All code will be released at https://github.com/ucam-eo/tessera.

View arXiv pageView PDFProject pageGitHub635Add to collection

Get this paper in your agent:

hf papers read 2607\.03949

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2607.03949 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2607.03949 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2607.03949 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

A Scaling Study for fMRI Foundation Models

arXiv cs.LG

This paper conducts a scaling study for fMRI foundation models, revealing that performance depends on the combination of pretraining data size, model size, and training duration, not just compute.

Scalable and Trustworthy Earth Observation Foundation Models

arXiv cs.LG

This chapter reviews design principles and current landscape of foundation models for Earth observation, highlighting the need for domain-specific adaptation, physically plausible representations, and consistent evaluation benchmarks. It includes case studies on harmful algal bloom prediction and adaptive monitoring station selection.

Predict before you train: Scaling Laws for particle physics foundation models

arXiv cs.AI

This paper demonstrates that scaling laws fit on small transformer models can accurately predict the loss of much larger models trained on particle physics jet data, enabling compute budgets to be translated into expected physics performance before large training runs. They release five pretrained models and the full training recipe.

AlphaEarth Foundations helps map our planet in unprecedented detail

Google DeepMind Blog

Google DeepMind introduces AlphaEarth Foundations, an AI model that integrates petabytes of Earth observation data into unified embeddings to map and monitor the planet at 10x10 meter resolution. The model's compact representations enable efficient planetary-scale analysis for applications in food security, deforestation tracking, and environmental monitoring.