Upper White flood-map completion

This tree scores every 30 meter cell in the Upper White River basin (HUC-8 05120201) for how much it looks like FEMA's current Special Flood Hazard Area, given terrain and distance-to-water. The score is P(sfha | hydro). TRI on-site releases for 2023 are an overlay, not the training target.

P(sfha | hydro) is a map-completion score, not a 100-year exceedance.

Calibrated OOF PR-AUC (0.36) beats the SFHA-rate baseline (0.10) and a HAND score (0.24). To three decimals: raw PR-AUC 0.369, calibrated 0.362, Brier 0.073 after isotonic calibration. Stage D samples p_sfha_calibrated.tif only.

Five Zone X plants have one wet cell in a 120 m window; site-mean P is 0.06 to 0.19; none clear 0.50 on the footprint.

THURSDAY POOLS is the only large-inventory terrain hit, and it is neighboring land.

June 2008 Appendix 2 does not cover the industrial core of 05120201.

Not a FIRM or site-level flood risk. TRI names are an overlay only.

Plant	2023 on-site release (lb)	Highest P in 120 m	Mean P (p_mean)	What the high cell is
THURSDAY POOLS	257,590	0.769	0.152	max off parcel, neighboring unshaded X, HAND = 0
FGF LLC	27,335	0.780	0.060	on-parcel floodway sliver; dry mean
ROYAL SPA CORP	4,950	0.774	0.113	waterbody cell off the office parcel
LINDE GAS & EQUIPMENT	1,048	0.763	0.192	max on the office parcel
MAGNA POWERTRAIN EAST	0	0.789	0.098	max on the office parcel

The table ranks on window-max P. Buffer-max is one 30 m cell; buffer-mean is the 120 m footprint.

• THURSDAY POOLS (p_mean 0.152) - max off parcel, neighboring unshaded X, HAND = 0.

• FGF LLC (p_mean 0.060) - floodway cell on the office parcel, so the lot is still not a wet footprint.

• ROYAL SPA CORP (p_mean 0.113) - waterbody cell off the office parcel.

• LINDE GAS & EQUIPMENT (p_mean 0.192) & MAGNA POWERTRAIN EAST (p_mean 0.098) - max on the office parcel; means stay below 0.50.

Figure 1. Basin disagreement on calibrated P(sfha | hydro).

• Dark red: Mapped SFHA and floodway.

• Cyan: Unshaded X with calibrated P >= 0.75 (same threshold as Table 1 window-max, pixel not plant). Map-completion on the FIRM, not a plant-level hazard list.

• Pale: Other interior.

Figure 2. Office point to window-max cell.

• THURSDAY POOLS (p_mean 0.152): Neighboring unshaded X, max off parcel.

• FGF LLC (p_mean 0.060): On-parcel floodway sliver.

• ROYAL SPA CORP (p_mean 0.113): Waterbody cell off the office parcel.

• LINDE GAS & EQUIPMENT (p_mean 0.192) & MAGNA POWERTRAIN EAST (p_mean 0.098): Max cell on the office parcel.

Note: Each panel title has p_mean. Wash is calibrated P; box is the 9x9 (120 m) window.

Figure 3. June 7-9, 2008 inundation (OFR 2008-1322) code 2 on the same HUC.

• Blue polygons: Martinsville and Paragon.

• Grey dots: 117 TRI office points (D2 stays 117 / 0).

Note: Appendix 2 is reach-scale; the industrial core is code 1.

Figure 4. Indiana 2025 parcels (GIS Data Harvest) on the five Table 1 sites only.

• Setup: Snap 30 m. D tables unchanged. Folder: logs/parcels/.

• THURSDAY POOLS (p_mean 0.152): Max cell off the office parcel.

• FGF LLC (p_mean 0.060): Floodway cell on the office parcel.

• ROYAL SPA CORP (p_mean 0.113): Waterbody cell off the office parcel.

• LINDE GAS & EQUIPMENT (p_mean 0.192) & MAGNA POWERTRAIN EAST (p_mean 0.098): Max cell on the office parcel.

Map (interactive): logs/stage_d/map.html. Interview note: docs/interview_note.pdf. Project: docs/PROJECT.md.

Summary

Three trees on one 30 m HAND grid (HUC-8 05120201). This repo is map completion.

Tree	Question
indiana_flood_completion	Which 30 m cells look like the current SFHA given terrain and distance-to-water?
white_river_stage_inundation	Which cells on one White River reach are wet at USGS 03351000 stage 11.00 ft and at 21.18 ft?
white_river_fim_compare	Does that HAND bathtub sit in the same neighborhood as USGS SIR 2011-5138 on the Nora window?

Maps writeup

101 of 117 in-HUC plants sit on mapped unshaded X. The other 16 are already floodway (2), SFHA (4), shaded X (8), or unmapped (2).

Limitations:

• Parcels: five Table 1 sites only (Indiana 2025 harvest, logs/parcels/). Rest of the HUC has no cadastral clip.
• No soil: hydrologic soil group is not in this model.
• Not a FIRM: P does not replace the effective flood map.
• 2008 is coverage: 117 plants in mask code 1, 0 in code 2. Appendix 2 reaches in the HUC are Martinsville and Paragon only.

Sibling occupancy study: indiana_hazmat_floodplain (parked, Stage 0 frozen 2026-08-25). This tree imports that occupancy as a freeze file. It does not recompute it.

File	Role
METHODOLOGY.md	Locked contract
AGENTS.md	Agent rules
CHECKLIST.md	Operator list
src/floodmap/	Claims, freeze, HUC, template, Stage 0
whiteforge/	GraphForge pin

Stage 0

python3.12 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
PYTHONPATH=src:. python3 scripts/run_stage0.py \
  --huc tests/fixtures/huc.geojson \
  --out logs/stage0_fixture
PYTHONPATH=src:. python3 -m pytest tests -q

Hard gate: freeze numbers match, claim scan clean, HUC non-empty, template EPSG:5070 at 30 m, product laws allow Stage 0. See METHODOLOGY.md.

Live WBD + NLCD 2021 (network; writes data/raw/ and data/interim/, gitignored):

PYTHONPATH=src:. python3 scripts/run_stage0_live.py

Stage A (network; live nlcd_2021 template required):

PYTHONPATH=src:. python3 scripts/run_stage_a.py

Stage B (network for NHD waterbodies; live template + DEM required):

PYTHONPATH=src:. python3 scripts/run_stage_b.py

Finite TWI on interior, toy-watershed tests in tests/test_hydro.py, bands on transform sha256 479ac37628bfd7e5. D1 is unshaded_x. OFR 2008-1322 Appendix 2 is a reach-scale three-state mask. HSG stays out of the B stack and out of this C model.

Stage C (HUC-10 fetch; live bands required):

PYTHONPATH=src:. python3 scripts/run_stage_c.py

Report and colorbar: P(sfha | hydro). Calibrate OOF scores before D:

PYTHONPATH=src:. python3 scripts/run_c_calibrate.py
PYTHONPATH=src:. python3 scripts/run_stage_d.py
PYTHONPATH=src:. python3 scripts/run_d_map.py
PYTHONPATH=src:. python3 scripts/run_d_shap.py
PYTHONPATH=src:. python3 scripts/run_d_cartography.py
PYTHONPATH=src:. python3 scripts/run_parcels.py

The fixture path is the CI join. gSSURGO C2 is not run. Do not reopen D, B, or raw p_sfha.tif.

Claim bans

Reports state hydrologic class and, later, TRI overlay tables. The scanner in floodmap.claims fails the run if the report text includes casualty language, climate attribution, tornado counts, population-at-risk, TRI storage as pounds, P as a 100-year exceedance, or the phrase “unmapped risk”.

GraphForge

Pin: whiteforge/. Verify-before-done is the finish gate.

Legal

Research index: https://gist.github.com/martialsystems/66b896b0a4a0b8cba2b478aef64312f3

Copyright (c) 2026 Martial Systems LLC. MIT. See LICENSE.