WEPPcloud

← Back to help

ADR-0054: Staley dNBR backend normalization

Date: 2026-09-09

Status and provenance

Accepted for backend implementation, 2026-09-09. Venue: user/Codex repository session. User requested the dNBR contract and WEPPpy implementation; Codex implements the bounded backend-first increment after stating that assumption. Scientific direction from prior user decisions: SBS format family, differing scales/extents, exact project alignment, and resilience to partial coverage. Browser/NoDb/RQ publication is deferred, not implicitly approved here.

Decision and rationale

Follow the backend contract. Accept real integer and float dNBR with explicit positive scale and finite offset. Normalize before compiled GDAL nearest sampling to Float32/NaN. Default metadata does not infer encoding; nondefault conflicting metadata fails. Negative/zero values survive. Nonfinite values are invalid support. Report rather than clip ideal-range departures. Source public fixtures are Float32 x1000, illustrating why dtype-based automatic scaling is unsound.

Nearest sampling preserves observations and source holes; bilinear/cubic interpolation can mix values or bridge holes. Report target-cell support, not exact source-polygon support. Partial means use observed equal-area target cells, and normalized mean supplies M1 F directly. No double /1000, implicit zero fill, minimum coverage cutoff, or effect on M3.

Alternatives and limits

SBS categorical class restrictions do not apply to continuous dNBR. A 30 m source does not become genuinely 10 m information by project-grid alignment. Unknown dates are allowed, but supplied date order is validated. Restrict VRT to explicit local identity wrappers and HFA/GTiff to self-contained inputs; full VRT execution/sidecar transport requires a later boundary contract. New completed artifact directories preserve prior results on failure without pretending to implement concurrent NoDb publication or invalidation.

Evidence, risk and rollback

See the active 20260908_staley_dnbr package for fixtures, kernel comparison, tests and independent review evidence. No predictive calibration claim. Nearest target support can miss tiny overlaps, which fail explicitly. Callers must validate WBT/locale readiness and own output publication. Removing this additive backend does not require a production-state migration.