Ingest: LANDFIRE 2024 CONUS — FBFM40 hex (finish 4 slices) + EVC decision

[cassiebuhler]

Split out of #590. Source: LANDFIRE LF 2024 Update v2.5.0, CONUS, bucket public-landfire.

State (verified against S3, 2026-09-16)

FBFM40 COG is built and on S3 (s3://public-landfire/landfire-2024-fbfm40/landfire-2024-fbfm40-cog.tif, 3.6 GB, WGS84, -r near). Hex has 2 of 6 h0 partitions. No stac-collection.json.

H3 native 10, parents 9,8,0 · reducer mode · id landfire-2024-fbfm40

h0-index	h0 cell	status
12	576812596024311807	present
14	577692205326532607	present
20	577164439745200127	missing
50	577199624117288959	missing
71	577762574070710271	missing
78	577234808489377791	missing

FBFM40 is now the only layer in the bucket with an incomplete hex. Ten others (vcc, evt, fbfm13, fvc, fvh, fvt, cbd, cbh, cc, ch) carry 6 of 6, including all four h0 cells missing here. evc and evh have COGs and no hex by decision.

What to do

Do not apply landfire-2024-hex-finish.yaml. It fans out 12 indexes (EVC and FBFM40 over 6 h0 each) at 192Gi with priorityClassName: opportunistic. The per-slice skip guard exits finished slices cheaply, but it still schedules 6 EVC pods that are not wanted and sizes the 4 FBFM40 pods from a superseded estimate.

Copy hex/fbfm13/landfire-2024-fbfm13-hex.yaml (PR #657) with --value-column fbfm40, scoped to the 4 missing h0. fbfm13 is the nearest analogue: same product family, same --nodata "-9999,32767", same res-10 mode. It completed 6/6 at cpu 4, memory 64Gi, CNG_HEX_WORKERS=8, default priority, parallelism 3, measured peak ~37 GiB. The earlier ~160Gi figure is superseded; over-requesting invites NRP's low-utilization admission rejection.

The two exit-137s are explained

Not a property of FBFM40. Slices run ~7 h and opportunistic is priority -2000000000, so a multi-hour pod is a large preemption target (.claude/skills/pod-preemption). PR #657 dropped opportunistic for exactly this reason, and its layers then completed on the same four h0 cells that failed here. The node co-tenancy hypothesis recorded earlier is not needed.

EVC

Settled to "not mode" by the #515 measurement: the modal class holds a minority of the cell in 63% of cells, median share 0.33. Two options remain, and both still need a stated consumer before building:

• decoded surface: (v − 100) / (v − 200) / (v − 300) into a percent column, lifeform as its own column, codes 11–100 as nodata, reducer mean
• fractions: preserves the lifeform-bin mix and makes area accounting exact

PR #657 defers EVC to this issue, so the decision stays here. EVH shares EVC's coded-bin structure and is deferred on the same grounds.

Blocked on

• PR feat(#515): LANDFIRE 2024 structure/fuels — 9 COGs + hex for the measured seven-layer build set #657 (open since 2026-09-07) carries the fbfm13 pattern and the 7 layers already on S3. The repo record lags the bucket until it merges.
• PR fix(#203): landfire STAC no longer advertises a fractions asset that was purged on purpose #655 (open since 2026-09-05) stops make-stac.py advertising a fractions asset that was purged on purpose. FBFM40's STAC inherits that defect if written first.
• Parent resolutions ship as 9,8,0 against 9,8,7,6,5,0 pinned on Import: LANDFIRE CONUS vegetation & departure suite — EVT, BpS, FRG, SClass, VCC, VDep (LF2024/2020/2016) #203 and Import: LANDFIRE 2024 CONUS vegetation structure & fuels — EVC, EVH, FBFM13/40, FVC/FVH/FVT, canopy CBH/CBD/CC/CH #515, with no recorded decision behind the difference. Applies to the whole tranche.

Caveats

• NB1–NB9 (91/92/93/98/99) are non-burnable real classes, not nodata: urban, snow/ice, agriculture, open water, bare rock. A hazardous-fuels statistic that does not split them out counts parking lots and lakes.
• Fill codes must not enter classification:classes (Fix: drop fill codes from classification:classes — landfire-2024-vcc + catalog convention #628).

Acceptance criteria

• All 6 FBFM40 h0 partitions present; job reports Complete=True with empty failedIndexes
• h0 coverage gate with --min-bytes (mode is sparse)
• NB1–NB9 documented as classes, excludable from fuels denominators
• stac-collection.json published and verify-stac.py clean
• EVC decision recorded in STAC, or a decoded-EVC issue opened

Part of #515 (#594 is closed). Evidence: catalog/landfire/k8s/landfire-2024/BUILD.md.

[cboettig]

Triage note: recorded the wider state of both LANDFIRE tranches on #203 and #515 today, since this issue is a child of #515 rather than a replacement for it.

Correction to an earlier version of this comment: I first wrote that the missing hex-fractions was an unexplained gap. It is not — BUILD.md records it as a deliberate, measured decision for #590's tranche (mode reproduced the true CONUS VCC distribution to within 0.72 pp; the partial prefixes were purged so they could not fake coverage). Details on #203.

That changes what it implies for the EVC question here, rather than removing it:

• The Ingest: LANDFIRE 2024 CONUS — VCC + EVT #590 rationale does not obviously transfer to EVC. It rests on cells being effectively pure — ~17 pixels at 30 m/res-10, stands far larger. EVC is the counter-case in this suite at 7.34 distinct values per res-10 cell (FBFM40: 2.14, max 16 of 16), which is exactly why mode was rejected for it here. So "cells are mostly pure, mode is close enough" is unavailable as an argument for EVC.
• That arguably makes fractions the more natural third option for EVC, alongside the two already on the table (record "COG published, no hex", or build a decoded surface). It sidesteps the objection in the caveat above — that mode over 7 near-identical cover percentages "picks a winner by coverage noise" — without needing the decode decided first, since fractions preserves the mix rather than choosing among it. Worth pricing before defaulting to "no hex": the scripts exist and landfire-2024-hex.yaml already carries the RED=fractions; OUT=hex-fractions branch.
• Parent resolutions are 9,8,0 as shipped against 9,8,7,6,5,0 as pinned on both umbrellas, with no recorded decision behind the difference. Whatever runs next here should use the settled answer.

Separately, a generator defect that will affect this issue's output: make-stac.py:182-188 builds the mode hex description unconditionally, including "for the area held by each class use the fractional-coverage asset instead", while the fractions_published=False branch only rewrites the collection-level description. Both published collections currently point at an asset that does not exist, and FBFM40 will inherit the same sentence when it publishes. Worth fixing in the generator before the next STAC write.

The FBFM40 half — finish the 4 missing slices at ~160 Gi — is unaffected by any of this and can proceed exactly as scoped.

[cboettig]

The EVC decision now has a measurement behind it — see #515

Measured whether mode is meaningful for these layers rather than inheriting #590's VCC result. Full method and caveats on #515; the part that settles this issue:

	vcc (control)	fbfm40	evc
distinct classes per cell	1.66	2.25	6.41
modal class share, median	1.00	0.85	0.33
cells where the winner holds < 50%	3.9%	12.8%	63.1%

Independently corroborates this issue's own 7.34 / 2.14 distinct-per-cell figures (6.41 / 2.25 here, different sample).

FBFM40 — proceed exactly as scoped. 78% mean modal share, 38% of cells pure, worst-class deviation 1.8 pp. mode is defensible for fuel models. Finish the 4 missing slices at ~160 Gi as planned; nothing here changes that half.

EVC — do not ship a mode hex. The caveat in this issue said mode over ~7 near-identical cover percentages "picks a winner by coverage noise". That is now quantified, and it is worse than "near-identical": in 63% of cells the winner holds a minority of the cell, median share one third. A mode value would describe a third of what is in the cell while presenting as a confident per-cell measurement.

Of the two options this issue lists, the measurement favours the decoded surface over "COG published, no hex":

(v − 100) / (v − 200) / (v − 300) into a percent column, lifeform as its own column, codes 11–100 as nodata, reducer mean

With a median modal share of 0.33, the categorical framing is what is doing the damage — EVC is quasi-continuous wearing categorical clothing, and mean over a decoded percent matches what the pixel actually measures. fractions is the complementary option: it preserves the lifeform-bin mix and makes area accounting exact. They answer different questions and are not exclusive.

This issue asked for a stated consumer before building the decode. That requirement still stands — the measurement says mode is the wrong answer, not that the decode is automatically wanted. But "record COG published, no hex in STAC" is now the weaker of the two fallbacks: it leaves the layer unusable for spatial joins when a usable representation is available at known cost.

Worth noting for #515 scope: EVH, FVC and FVH share EVC's exact coded-bin structure and are untested. Expect EVC's behaviour rather than FBFM40's, and measure before building — the job is ~90 s per layer at 2 cpu / 8 Gi with no localization.