Results

Products of the real processing chain, with per-band quality statistics and their reading. Two validated end-to-end runs back this page; both consume the Sentinel-2 MSI Synthetic Raw Data Generator’s products as input.

L0 → L1B — real chain, persisted product

The generator’s open-container L0 + calibration-database ADFs pushed through l0_decode → radiometric → enhancement → toa (eopf == 2.8.1, nominal mode, emit_reflectance): a persisted L1B TOA-reflectance EOPF product.

L1B TOA reflectance, RGB = B04/B03/B02, per-channel 2–98 % percentile stretch. The demo scene is a flat field, so the stretch deliberately reveals the residual texture: per-column striping is the impressed PRNU pattern, the speckle is shot/read noise.

Per-band statistics (the pipeline’s stats phase, the non-referential ALG-QA metrics of msi_processor/common/metrics.py — SDD <5.4.1>; produced by the manual product-stats CI job, 2026-07-02):

Band	mean (reflectance)	std	variance	SNR (dB)
B02	0.1753	0.0053	0.000029	30.3
B03	0.1888	0.0066	0.000044	29.1
B04	0.1911	0.0070	0.000049	28.7
B08	0.2648	0.0094	0.000089	29.0
B11	0.0434	0.0017	0.000003	28.3
B12	0.0535	0.0019	0.000004	28.8

Reading the numbers

• Reflectance means sit at the levels impressed by the generator’s flat-field scene (≈ 0.19 VNIR / 0.27 NIR / 0.05 SWIR) — the absolute radiometric scale survives the full chain (DN → radiance → reflectance) unchanged.

• SNR is scene-limited, not instrument-limited: on a flat field the standard deviation is the residual instrument texture (uncorrected PRNU + noise), so the per-band SNR ranks the bands by their noise-model α/β and PRNU amplitude.

• Variance complements SNR as the raw second moment used by the QA stage (ALG-QA-*); both are computed product-side with no reference.

Validation status

• The chain producing these numbers is CI-verified: 8 of 8 processing units implemented and green (see the SUITR and the live CI test report).

• A second, sharper end-to-end result — the real-L1A bit-identity run through l0_decode (L1A′ ≡ L1A, 13/13 bands) — is documented in the generator’s validation pages and its qualification is summarised in the QR report.

• AR-gated figures: the referential accuracy metrics (L1B radiometric RMSE vs a calibrated reference — REQ-P-01 RAD_ACC; geometric GEO_CE90 / BAND_COREG — REQ-P-02; BOA_ACC — REQ-P-03) are validated on operator data at AR and are deliberately not quoted here (see the V&V report). the stats phase computes them once a reference product is available.

Reproduce

Run the manual pipeline-nominal CI job (integration-tests stage): the repository’s single driver pulls the inputs from the shared ipf/data-store registry, runs l0-decode → radiometric → enhancement → toa → stats and artifacts the statistics table. Locally (eopf environment): python scripts/run_pipeline.py <store>; --mode calibration derives and cross-validates the NUC instead.