hazus-curves

An open, tidy database of FEMA Hazus damage and vulnerability curves — flood and hurricane wind — that installs into a local SQL environment in one command.

Hazus is the most widely used natural-hazard loss model in US practice, but its damage functions live inside a Windows-only ArcGIS Pro application backed by SQL Server. Getting at them normally means a ~50 GB install and a licensed ArcGIS Pro seat, or copying tables out of PDFs by hand.

This project extracts them, reshapes them into tidy long format, documents where every number came from, and packages the result so a researcher can start querying in under a minute.

→ Browse the curves in your browser — no install, no account. Filter, chart and download curves as CSV or JSON.

Not affiliated with or endorsed by FEMA. "Hazus" is a trademark of the Federal Emergency Management Agency. See LICENSE-DATA.md.

Not on PyPI yet. pip install hazus-curves does not work today — the first release has not been published. Install from GitHub instead; everything below behaves identically either way.

pip install "git+https://github.com/rkvaughn/hazus-curves@v0.1.0"   # until PyPI is live
hazus-curves install                 # flood curves -> ~/.hazus_curves/ (7 MB)
hazus-curves query "SELECT curve_id, occupancy, source_agency FROM curves LIMIT 5"

Once published this becomes:

pip install hazus-curves

from hazus_curves import connect, get_curve, damage

con = connect()
damage(con, "fl-6.1-structure-129", 3.5)     # % structure damage at 3.5 ft depth
curve = get_curve(con, "fl-6.1-structure-129")
curve["kind"]["y_units"]                      # 'percent' - never assume units

Other engines, from the same artifacts:

hazus-curves install --perils fl,hu --target duckdb:///hazus.duckdb
hazus-curves install --perils fl    --target postgresql://user@localhost/hazus

DDL for SQLite, DuckDB, PostgreSQL and Snowflake is generated from a single schema definition (hazus_curves/schema.py) into sql/, so the engines cannot drift apart.

Flood curves are depth-damage functions: depth in feet relative to the first finished floor (−4 to +24 ft, 29 points), damage as percent of replacement value, split into structure, contents and business inventory, keyed by occupancy class and flood zone.

Hurricane curves run over 3-second peak gust wind speed (50–250 mph in 5 mph steps), for 6,116 wind building types × 5 terrain roughnesses × 9 loss classes. The nine loss classes include four damage-state exceedance probabilities, building and content loss ratios, loss of use in days, and two debris measures in lbs/sq ft.

Those nine classes do not share units. Join the curve_kind table rather than assuming — see docs/schema.md.

No fabricated values. The pipeline renames columns and reshapes wide tables into long ones. It never computes, interpolates, smooths, clamps, or infers a damage value. Gaps stay gaps — a missing source value becomes NULL, never an estimate. Every row carries source_file, source_table and source_row_id back to the originating cell, every source file is pinned by SHA-256, and the round-trip is enforced by tests.

That rule is why several things below are reported as problems rather than quietly fixed.

Three things came out of building this that are not documented elsewhere.

FEMA's release notes say the flood library was expanded but not whether existing curves were revised. Measured directly:

Every curve that existed in Hazus 4.0 survives into 6.1 bit for bit, so results computed against the older library remain reproducible. → docs/version_changes.md

FEMA's Hazus 7.1 release notes disclose that in 6.1 and 7.0 the 2- and 3-story multi-family wind curves had been overwritten with copies of the 1-story curves. We tested that claim against the data instead of taking it on faith, comparing every 2-/3-story curve with the 1-story curve having identical characteristics, terrain and loss class:

15,200 of 34,560 multi-story multi-family curves are byte-identical to their 1-story counterpart.

The defect is real and measurable. Affected curves ship with a defect_flag column in the data, not a footnote, plus a defect_verified column recording our own measurement separately from FEMA's disclosure. We flag rather than withhold — a flagged curve is more useful than a missing one, and Hazus 7.1+ (which has the fix) is not publicly extractable. → docs/hurricane_defect.md

The Hazus 6.1 workbooks reached us via a public bucket belonging to os-climate/physrisk, not directly from FEMA, and how they obtained them is undocumented. That is a real provenance gap — so we tested it rather than assert it.

• Flood: the 6.1 workbook embeds the pre-6.1 structure library; against FEMA's own published Hazus 4.0 CSVs, all 597 curves across all 29 depth values are identical, zero discrepancies.
• Hurricane: matches the FEMA Technical Manual on four independent counts — 62 wind building characteristics, 39 specific building types, "over 275,000 damage functions" (275,220), and it reproduces a specific defect FEMA disclosed in its 7.1 release notes (15,200 confirmed duplicate curves). A re-derived dataset would not carry that.

⚠️ That bucket went dark on 2026-08-12 (403 AllAccessDisabled, whole bucket, no Internet Archive capture). The originals can no longer be downloaded by anyone, so re-downloading and byte-comparing is no longer possible. We re-host our SHA-256-verified copies on the v0.1.0 release; fetch.py falls back to them and rejects any file whose hash does not match the committed manifest.

→ docs/provenance.md

Also worth knowing before you use the data: some upstream probabilities exceed 1.0, and nine building-characteristic codes are missing from Hazus's own decode table. Both are preserved as-is and documented in docs/data_quality.md.

Everything here is checkable against FEMA's own documents, and the verification report (Word, 23 pages) does exactly that: nine quantities FEMA states in four published documents, each tested against this dataset, with the source page reproduced as a screenshot alongside its URL and page number. Eight match; one is a partial match and is reported as a discrepancy rather than presented as agreement.

All eleven FEMA manuals and release notes it cites are committed under raw/, so the evidence base travels with the repository. Regenerate the report with:

.venv/bin/python scripts/build_verification_evidence.py   # locate, render, measure
node scripts/build_verification_docx.js                   # build the document

python -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/python scripts/build_all.py --perils fl,hu   # fetch + build everything
.venv/bin/python -m pytest                             # 114 tests

The repository carries the FEMA source documents (~101 MB), so a full clone is correspondingly large. If you only want the code:

git clone --depth 1 --single-branch https://github.com/rkvaughn/hazus-curves.git

The Parquet tables behind the web tool live in a separate repository, hazus-curves-data, so they are not pulled by a clone or by pip install. That repository documents itself — schema reference, worked queries and provenance — and can be used without this one. Its documentation is generated from the schema and the data:

.venv/bin/python scripts/build_data_repo_docs.py --out ../hazus-curves-data

build_all.py runs the whole pipeline; each stage is also runnable on its own (fetch.py, build_flood.py, build_hurricane.py, verify_hurricane_defect.py, extract_building_types.py, diff_versions.py, build_db.py).

fetch.py is idempotent: re-running verifies checksums rather than re-downloading.

• Vintage is Hazus 6.1, the newest publicly extractable release. Flood curve values are unchanged through Hazus 7.1, so flood is effectively current; hurricane is not, and carries the defect above. Hazus 7.2's release notes ship inside a CAPTCHA-gated download, so what changed there is unknown.
• No earthquake or tsunami yet. Their parameters are in technical-manual PDFs. See the roadmap.
• No geography. The Hazus curve tables contain no state or county key, so this database cannot map an address to a curve. What it does carry is flood zone and source region (USACE district, FIA, FEMA PFRA). An open address-selection model is on the roadmap as a separate project.
• The flood curves substantially overlap existing prior art. The value added here is the schema, the version diff and the packaging — not that the data was unavailable. Hurricane wind in open tidy form does appear to be new. See docs/prior_art.md.

Code: MIT (LICENSE). Data: believed to be a US Government work under 17 U.S.C. § 105, with the uncertainties stated plainly in LICENSE-DATA.md.

Please cite FEMA and the originating agencies — see CITATION.cff.