Geographic split audit

Geographic split audit, 2026-10-20

Canonical source: versioned documentation. This wiki copy is maintained manually.

The complete imported PV4GER-SEG release was split using geographic-block-buffer-v1, seed 20261020, 10 km metric cells and a 1 km buffer. The seed and policy were fixed before any model training or evaluation. The machine-readable report contains the fingerprints, source coordinate hash and policy parameters. No imagery or per-sample labels are included in this repository.

Partition or exclusion	Samples
Training	3,282
Validation	278
Test	421
Unassigned: geographic buffer	38
Unassigned: upstream zero mask pending review	8
Total	4,027

The policy uses 3,628 linked source components and 360 geographic blocks. The minimum measured centre distances are:

Pair	Metres, EPSG:25832
Training ↔ validation	1,012.82
Training ↔ test	1,010.49
Validation ↔ test	1,146.68

The final snapshot validates with no errors. Its 12,081 warnings preserve the known missing acquisition dates, missing original georeferences and upstream annotation status. Source pixels are unchanged. Repeating the split with the same parameters produced the same canonical dataset fingerprint: 597b9facb7f115448e81af3b88de7e1081d510c38108c294195de05763d841c5.

This is a buffered internal holdout from one source. It is not a new independent benchmark, does not exclude all possible city/campaign correlations, and cannot prove independence from existing pretrained models. Local label review and additional reviewed hard negatives remain outstanding. No model performance has been measured on this test partition.