ra1nbowdash/sentinel2-matching-pairs

Sentinel-2 Matching Pairs (sample)

Submission-sized sample of multi-temporal / seasonal satellite image pairs for the DS internship Task 2 (image matching).

Field	Value
Sample pairs in this repo	20
Full local dataset	920 pairs
Sources in sample	api_crops (8), oscd (6), ukraine (6)

Layout

pairs.jsonl          # index
manifest.json        # counts + notes
api_crops/.../t0.png # seasonal Ukraine Sentinel Hub crops
api_crops/.../t1.png
oscd/<city>/t{0,1}.png
ukraine/.../t{0,1}.png

Each pairs.jsonl row has path_t0, path_t1, source, split, and seasonal.

Data sources

This sample combines three upstream sources:

Source key	Description	Pairs (full / sample)
api_crops	Seasonal Sentinel Hub crops over Ukraine AOIs (1024×1024 windows)	864 / 8
ukraine	Multi-date pairs from the Kaggle Deforestation in Ukraine dataset (Isaienkov, 2023) — earliest vs latest date per window	32 / 6
oscd	Onera Satellite Change Detection (OSCD) city pairs	24 / 6

The ukraine/ folder in this repo holds pairs derived from Deforestation in Ukraine. Please respect the original Kaggle dataset license and terms when reusing those images.

Model weights

This assessment uses pretrained LoFTR via Kornia (kornia.feature.LoFTR), not a custom-trained checkpoint. Classical matchers (SIFT / ORB) need no weights.

After installing kornia, weights download automatically on first LoFTR run. Document the Kornia / LoFTR upstream release you use in your task README.

Full dataset

The complete 920-pair index lives in the local project under Task_2_CV/data/matching_dataset/ (large; not fully uploaded here). Rebuild scripts: Task_2_CV/scripts/build_matching_dataset.py, fuse_matching_dataset.py.

License note

Sentinel-2 / OSCD / Kaggle Deforestation in Ukraine source terms still apply to the imagery. This card only packages a sample for assessment reproducibility.