NJ Hazard Vulnerability Dashboard

Problem

The NJ Parcel Flood Risk Dashboard answers an underwriter’s question: how exposed is this property, today, on one number. Planners ask a different one. They need to know how the answer changes — which parcels screen clean under today’s FEMA map and badly under sea-level rise, where a Category 3 surge reaches that a Category 1 does not, and whether the places that come out worst are also the places least able to absorb the loss.

That is not a variation on the same product. Same parcel backbone, different score, different audience: state, county, and municipal resilience staff comparing exposure across scenarios, not underwriters pricing a single policy.

Approach

Every parcel is scored once per scenario on a state-resilience-aligned framework rather than a single-hazard finance score:

V = round(100 × (0.45·E + 0.30·S + 0.25·C))

• Exposure (E) — how much of the parcel intersects the hazard zone under this scenario, with a presence floor and a small bonus for parcels exposed across three or more scenarios at once.
• Sensitivity (S) — how consequential exposure is for this parcel’s use: locked class weights (critical and community-serving 1.00, residential 0.85, commercial 0.60, vacant 0.15) plus its improvement-to-total value ratio.
• Consequence context (C) — CDC social vulnerability percentile, NJDEP Overburdened Community status, and tax-base concentration.

The five scenarios are current FEMA flood at the 1% and 0.2% annual chance (S1, S2), sea-level rise of five feet (S3), and SLOSH hurricane storm surge at Categories 1 and 3 (S5, S6).

Results

Scenario	Parcels screened	% exposed	Mean V	Assessed value exposed
S1 — Current SFHA (1%)	3,477,980	13.0%	34.8	$286.7B
S2 — Current 0.2% shaded X	3,477,980	5.2%	31.8	$127.3B
S3 — SLR +5 ft (coastal)	2,852,337	18.7%	37.8	$306.3B
S5 — Storm surge Cat 1	3,117,435	10.9%	34.9	$216.2B
S6 — Storm surge Cat 3	3,117,435	19.7%	38.6	$330.1B

The spread is the finding. A parcel screening clean under today’s FEMA map can screen at real risk under sea-level rise or a Category 3 surge: S3 and S6 each expose roughly 50% more of the state’s assessed value than S1 does, precisely because they model conditions beyond what current infrastructure and flood control protect against. A dashboard that showed only S1 would be quietly reassuring and wrong.

Calibration, and the miss that stayed in

Nine municipalities with well-documented hazard history must rank top-quartile under S1 and S3. 14 of 18 checks pass. The other four are the interesting part:

• Three are correctly not screened. Bound Brook, Manville, and Wayne sit in non-coastal counties that S3 has no data footprint for at all. The distinction between “no data here” and “safe here” is load-bearing, and the dashboard never collapses it.
• One is a genuine miss. Ventnor under S1. Investigated rather than waved through: its built parcels really do sit outside FEMA’s current mapped flood zone, confirmed against the live FEMA service directly. The same town passes cleanly under S3, at 99.9% exposed, once sea-level rise is modelled past existing dune protection. Documented, not tuned away — the rule against adjusting weights to force a calibration pass is absolute.

Technical detail — pipeline scale, and four bugs worth naming

Five scenarios across 3,477,980 parcels is 17,389,900 fact rows, published as 585 summary JSON files, a 518.1 MB parcel tileset, a 2.1 MB boundaries tileset, and 564 search shards. The heaviest stage — the parcel × hazard × scenario intersection — ran 18.79 hours wall-clock, parallelised across five worker processes. Per-county cost tracks hazard-layer vertex density far more than parcel count: Cape May’s 143,206 parcels took 6.58 hours against Ocean County’s 421,548 in 13.30. A first statewide attempt was killed partway through by an unrelated session teardown and resumed cleanly from county checkpoints — no completed county was lost, only wall-clock time.

Four bugs are worth naming because of how each was caught:

• A statistical bug caught by an independent gate, not by inspection. The statewide KPI rollup understated % exposed for every partial-coverage scenario by 10–18%, because the summary function’s “coverage is uniform within a group” assumption — true at county and municipality grain by construction — silently broke at state grain, where counties differ in scenario coverage. Caught by a rollup-invariant gate that re-derives the state total from the published county files rather than trusting either number.
• A bug the aggregate count caught and the individual gate missed. A QA gate’s own bounds check used a pandas .all() chain, which returns numpy.bool_ rather than a Python bool. It printed PASS per-gate and silently failed the summary’s identity check, undercounting passes with no gate ever marked FAIL. Caught because the total did not match the number of PASS lines printed.
• A privacy-audit false positive, verified rather than assumed. A scan for owner-identity fields flagged seven search-index files. All false positives: “Downer Way,” “Pitman Downer Rd,” and “Downer Avenue” all contain the substring owner. Confirmed by inspecting every flagged file before concluding it was a false alarm, then fixed with word-boundary anchoring.
• A suspected upstream bug, correctly ruled out. A suspected ~80% completeness gap in the sibling project’s flood-zone data turned out to be a bounding-box mistake in the investigation itself, confirmed by re-testing against the live FEMA service with the correct county boundary. The false lead is documented so a later session does not repeat it.

A genuine upstream artifact did survive all this: 742 duplicate MOD-IV PINs statewide (0.02%), which crashed the fact table’s grain-uniqueness assertion on the first attempt and turned out to explain a previously unexplained 742-parcel discrepancy against the shared parcel core’s canonical count.

Architecture

The parcel core is shared with the sibling project, read-only. Everything downstream is this project’s own: hazard-layer normalisation, the scenario intersection, E/S/C scoring, summaries and calibration, tiles, and 11 QA gates — all re-derived fresh from published data rather than trusted from any earlier stage’s self-report.

The web app is React 19, Vite, Tailwind v4, and MapLibre across four views, and it is deliberately differentiated from its sibling in the interface as well as the analysis: a left sidebar with the scenario selector as a persistent control rather than a top nav bar, warm amber chrome against a cool colourblind-safe severity ramp (the sibling inverts both), sharper corners, a different icon mark. Two dashboards on one backbone should not be mistakable for each other at a glance.

Hosting is static — Cloudflare Pages and R2, the same zero-backend shape as the sibling.

Data sources

Dataset	Source
Parcel core	NJ MOD-IV, shared with the NJ Parcel Flood Risk Dashboard
Current flood zones	FEMA National Flood Hazard Layer
Sea-level rise	NJDEP CAFE
Storm surge	NOAA SLOSH MOM, via Rutgers
Social vulnerability	CDC/ATSDR SVI 2022
Environmental justice	NJDEP Overburdened Communities

Screening tool, not a determination. Results are for hazard screening and prioritisation research; parcels must not be targeted for action without full analytical and policy review.