Deficit-Driven Stratified Acquisition for Building STRATA-S12: An Open SAR–Optical Corpus for Self-Supervised Learning
Christian Sanchez-Leon
PhD student, Northeastern University, Shenyang
Research
I work on self-supervised learning for Earth observation using paired Sentinel-1 SAR and Sentinel-2 optical imagery. A large part of my work is on the data side: how to build open pretraining corpora whose geographic and land-cover coverage is controlled by design rather than inherited from whatever imagery is easiest to download.
Before the PhD I worked as a software engineer, and I try to bring the same habits to research: pinned versions, checksums for every artifact, and pipelines that fail loudly instead of producing quiet nonsense.
Publications
Data and code
- STRATA-S12
-
740,746 co-registered Sentinel-1 / Sentinel-2 pairs over six continents,
sampled across 192 strata. CC BY 4.0.
Shards on Hugging Face - STRATA-S12 corpus (D) and acquisition ledgers
-
Citable record of the full corpus (about 2.7 TB, data hosted on Hugging Face),
with the public records of rejected and withdrawn pairs.
doi:10.5281/zenodo.22699046 - STRATA-S12 Lite (D_lite)
-
A compact, class-balanced subset of 20,916 pairs, for working with the
data directly without downloading the full multi-terabyte corpus.
doi:10.5281/zenodo.22646276 - strata-s12-eval-harness
-
Linear-probe evaluation of pretrained encoders on reBEN (BigEarthNet v2.0).
GitHub doi:10.5281/zenodo.22892250