TESSERA Produces Pixel-wise Global Earth Embeddings

TESSERA, released with open weights, code, and precomputed global embeddings, compresses a year of Sentinel-1 and Sentinel-2 time series into 128-dimensional pixel-wise representations at 10m resolution. The project publishes global, annual, int8 embeddings so teams can skip expensive time-series preprocessing and run retrieval or inference at planetary scale.

`TESSERA` is a pixel-oriented foundation model for multi-modal Earth observation that uses self-supervised learning to produce temporally aware embeddings. Training incorporates Barlow Twins and a sparse random temporal sampling strategy to enforce invariance across irregular observation sequences caused by orbit geometry and cloud cover. Two additional regularizers are introduced to improve robustness:

• •global shuffling, to decorrelate spatial neighborhoods and reduce overfitting to local cues
• •mix-based regulation, to boost invariance under extreme observation sparsity

The published artifact is a 128-dimensional latent vector per 10m pixel, quantized to int8, paired with lightweight adaptation heads for common downstream tasks. The project ships a user-facing Python package, `geotessera`, plus example pipelines demonstrating pixel-level segmentation, classification, and regression.

Benchmarks and performance TESSERA embeddings deliver state-of-the-art accuracy and strong label efficiency across diverse tasks, often requiring only a small task head and minimal labeled data. The team reports improvements across classification, segmentation, and regression benchmarks, and collaborators have applied the embeddings to concrete problems such as mapping solar farms with compact models and semantic segmentation at fine spatial scale.

This release is part of a larger movement toward Earth-observation foundation models and precomputed planetary embeddings. By publishing global, precomputed embeddings, TESSERA follows a pragmatic pattern: move heavy compute and time-series modeling into a one-time offline build, then serve compact artifacts that unlock low-cost downstream workflows. The approach lowers entry costs for teams without GPU farms, accelerates iteration, and enables near-real-time operational uses that would otherwise require expensive on-the-fly time-series processing.

Limitations and caveats The current model focuses on Sentinel-1/2 sources and annual aggregation, which suits many ecological and land-cover tasks but may miss sub-annual events or require augmentation for higher-resolution or hyperspectral needs. The int8 quantization optimizes storage and serving at planetary scale, but it reduces representational fidelity compared to full precision. As with any foundation model, regional biases, label distribution shift, and temporal transfer remain open issues to monitor in production deployments.

Adoption of the `geotessera` tooling and community-driven fine-tuning will determine whether precomputed embeddings become a standard interface for EO analytics. Expect follow-ups that expand modalities, increase temporal granularity, and provide integration layers for geospatial-aware LLMs and downstream model zoos.