// the take
TESSERA is a self-supervised foundation model that turns a year of Sentinel-1/2 time series into per-pixel 128-d embeddings at 10m global resolution, trained with a Barlow Twins objective on cloud-corrupted satellite data. It's aimed at remote sensing researchers doing land cover, crop type, or canopy height work who want off-the-shelf embeddings instead of training task-specific models from raw imagery.
The v2 students emit Matryoshka embeddings (16/32/64/128-d truncatable from one checkpoint, no retraining needed) which is a genuinely useful space/accuracy tradeoff for downstream storage. Most users never need to run the pipeline at all — precomputed global embeddings are downloadable via the separate GeoTessera package, and this repo is only needed if you want custom regions/years. The README is unusually explicit about the sharp edges that would otherwise silently corrupt results: the non-standard Sentinel-2 band order tied to the checkpoint weights, and the fact the 2B teacher and distilled students use different S1 ascending/descending normalization schemes.
There are four non-interchangeable inference codepaths (v1.0 early, v1.0 QAT, v1.1, v2), each with its own checkpoints and normalization stats — the README even documents a past incident where the AWS preprocessing pipeline double-applied a BOA offset and silently degraded a released checkpoint, which is the exact failure mode this fragmentation invites. Running your own inference is a heavyweight ask: Linux-only preprocessing (no macOS), 1TB+ storage and 128GB RAM for a modest 100km² tile, plus hand-editing bash scripts and chmod'ing shell wrappers rather than a proper CLI or pip package. The repo also carries three near-duplicate profiler directories under each inference variant, which is dead weight for anyone just trying to run inference.
We read the README, the directory tree and a key source file in September 2026. Nobody pays to be listed. How we judge →