Flood-Cascade: following flood-mobilised contamination downstream

Framing (non-negotiable). The unit of analysis is the US county (5-digit FIPS), built up from S2 Level-13 grid cells (~1.2 km) and NHDPlus stream reaches. Coverage is the modelled UF-OKN flood footprint (national, 47,512 buildings) intersected with the OKN federation's hydrologic network (507,486 NHDPlus reaches across 348 HUC8 subbasins in 10 HUC2 regions). The level of inference is spatial-topological plausibility: a routed link means water demonstrably flows from A to B, not that a contaminant was released, transported, or measured. This is scenario-based hypothesis generation from model output and regulatory inventories — not field measurement and not causal inference. Keep that caveat attached to every downstream claim in this report.

Abbreviations. COMID = NHDPlus common identifier for a stream reach; EPA FRS = US Environmental Protection Agency Facility Registry Service; FIPS = Federal Information Processing Standard county code; HHI = Herfindahl–Hirschman Index; HUC = hydrologic unit code (HUC2 = region, HUC4 = subregion, HUC8 = subbasin); KG = knowledge graph; NAICS = North American Industry Classification System; NHDPlus = National Hydrography Dataset Plus (v2); NPDES = National Pollutant Discharge Elimination System; PFAS = per- and polyfluoroalkyl substances; PWS = public water system; RCRA = Resource Conservation and Recovery Act; RUCC = Rural–Urban Continuum Code (1 = metro core … 9 = most rural); S2 = Google S2 geometry grid; SDWIS = Safe Drinking Water Information System; UF-OKN = Urban Flooding Open Knowledge Network; WQP = Water Quality Portal.

1. Executive summary

A flood is a transport event, not a static exposure. Existing burden maps — including the OKN federation's own environmental-justice work — score a county on what sits inside its own boundary. This study asks the question that framing cannot: when the water leaves, where does it go, and who is standing there?

Starting from the UF-OKN modelled flood footprint (47,512 flood-impacted buildings resolving to 2,738 S2 Level-13 cells nationally), we found 11,085 EPA-regulated facilities sitting inside 1,339 of those cells — a co-location rate of about 54% of the flood cells that could be placed on the national grid. Flood exposure of regulated industry is moderately concentrated, not diffuse: of the 2,270 facilities carrying a NAICS code, manufacturing (NAICS 31–33) alone accounts for 799 (35.2%), the top three sectors take 45.1%, and the sector Herfindahl–Hirschman Index is 908.0.

We then routed the water. Using hydrologykg's precomputed NHDPlus transitive downstream closure, 1,101 stream reaches crossing a facility-bearing flood cell reach 15,727 distinct downstream reaches over 996,529 routed links, touching 25,063 downstream S2 cells. Rolled up to counties, this produces a typology rather than a ranking: 153 Retained counties (their contamination stays home), 149 Imported counties (their exposure is generated upstream), 59 Compound counties (heavy on both), and 158 Low. 179 counties, home to 18.5 million people, have no flood-exposed regulated facility of their own yet sit downstream of one. They are invisible to any co-location metric.

The imported-risk group is the point of the exercise, and it has a distinct social signature. Imported counties are markedly more rural and smaller than Retained ones (62% vs 33% rural; median RUCC 6 vs 3; median population 20,813 vs 107,215). Dropping the routing step and ranking on co-location alone changes 31 of the top 50 counties (Spearman ρ = 0.578, Kendall τ = 0.475). And 151 of the 149+59 Imported/Compound counties — 8.4 million residents — have no contaminant monitoring feature anywhere in their downstream cells in the federation. What this adds is a reproducible, federation-native method for turning a static burden map into a directed one.

2. Sources used

Every row above traces to a logged, non-exploratory SPARQL query in the reproducibility record. geoconnex was examined during design (its downstreamWaterbody mainstem layer and county predicate) but contributes no logged result to any finding and is therefore deliberately absent from this table; the routing is carried entirely by hydrologykg, whose reach subjects are minted as geoconnex COMID IRIs.

3. Design & rules

The design is a four-stage chain, each stage joining on a geographic key, never a name.

Stage 1 — the flood footprint. UF-OKN publishes flood-depth predictions as schema:Observation records about individual buildings, with a qudt Depth property value in metres and the building's coordinates on a schema:GeoCoordinates node. We took every building with a depth prediction (47,512 distinct buildings; median modelled maximum depth 1.94 m) and computed its S2 Level-13 cell client-side, verified against the server's own point_to_s2 conversion. That yields 2,738 flood cells, of which 2,503 fall inside spatialkg's contiguous-US grid and resolve to 358 counties.

Stage 2 — what sits in the footprint. For each flood cell we asked fiokg for every EPA FRS facility whose sfWithin S2 cell is that cell, collecting its NAICS codes (all hierarchy levels) and its environmental-programme interests. This is deliberately cell-resolution co-location, not county-level: county-level co-location is the naive baseline this study exists to beat.

Stage 3 — routing. hydrologykg mints each NHDPlus reach as its geoconnex COMID IRI and carries two things that make routing possible: sfCrosses links from a reach to every S2 cell it passes through, and downstreamFlowPathTC, a precomputed transitive downstream closure. A reach is a source reach if it crosses a flood cell that contains at least one facility. From each source reach we take the full downstream closure, map every downstream reach back to its S2 cells and thence to counties, and label each source→downstream link by hydrologic proximity (§4).

Stage 4 — who is there. Receiving counties are profiled on rurality and population (ruralkg RUCC), on whether any contaminant-monitoring feature exists in their downstream cells (sawgraph), and on the flooded-well direct pathway (hydrologykg, §6.2).

Two things are visible immediately. The flood footprint is genuinely national (26.1°N–49.3°N, 122.7°W–68.3°W) while the routable network is not, and the facility-bearing subset of flood cells is concentrated in exactly the industrial river corridors — the Ohio, the Upper Mississippi, the Great Lakes shore — where the routing question matters most.

4. Confidence tiers

Every downstream claim is graded by how strong the hydrologic connection behind it is, using the hydrologic unit hierarchy as a distance proxy (the transitive closure gives connectivity but not path length, so HUC nesting is the honest available surrogate).

Tier C dominates by count (370,004 links), which is exactly what an unbounded transitive closure on a mainstem network produces and precisely why the tier weighting exists: a county 900 km down the Ohio is connected to a flooded plating shop in Illinois, but that connection should not carry the same weight as one 15 km downstream in the same subbasin. All imported-risk scores in §5 are tier-weighted; the unweighted counts are also reported in the workbook so a reader can see how much the weighting changes the picture. The absence of any Tier D link is a useful negative control: the routing never crosses a continental divide, as it should not.

5. Findings by axis

5.1 Which industries are flood-exposed, and how concentrated is that exposure

Of 11,085 flood-exposed EPA-regulated facilities, 2,270 carry a NAICS classification in fiokg (a 20% coding rate that is itself a finding — see §10, limitation 4). Within the coded set, exposure is moderately concentrated: manufacturing sectors 33, 32 and 31 together contribute 799 facilities (35.2%), and the three largest sectors take 45.1% of the total. The sector-level HHI of 908.0 sits in the range conventionally read as moderate concentration — flood exposure is not spread evenly across the economy, but neither is it a single-industry story.

The most flood-exposed six-digit industries are revealing in their ordinariness: general automotive repair (811111, 67 facilities), automotive body and paint shops (811121, 61), dry-cleaning and laundry services (812320, 55), electrical contractors (237310, 44), natural-gas distribution (221320, 33), and — the classic PFAS and metals concern — electroplating and anodising (332813, 22). These are small, widely distributed, chemically active premises, not a handful of mega-sites. Reading the EPA programme interests rather than NAICS gives the same picture from the regulatory side: 1,159 flood-exposed facilities are in a hazardous-waste programme, 1,220 are conditionally exempt small-quantity generators, 594 hold underground storage tanks, 677 hold minor NPDES discharge permits, 294 are Toxics Release Inventory reporters, 254 are brownfield properties, and 12 carry a Superfund or Superfund-non-NPL interest.

The takeaway is that a flood-contamination programme aimed only at large NPL sites would miss most of what is actually in the water's way; the exposed population of facilities is dominated by small solvent-, fuel- and metal-handling premises whose individual risk is low but whose count is high.

5.2 Where does the water go

The routing produces 996,529 source→downstream reach links from 1,101 source reaches — a mean fan-out of 905 downstream reaches per source, with a long right tail. Those reaches touch 25,063 S2 cells. Rolled to counties and restricted to genuinely imported links (downstream county ≠ source county), 266 counties receive at least one routed connection, and the distribution of upstream contributing counties is heavily skewed: most receiving counties draw from a handful of upstream counties, while the confluence counties at the bottom of the Ohio and Upper Mississippi draw from as many as 62.

Panel B is the honest picture of what an unbounded transitive closure buys you: enormous reach, and therefore the necessity of the tier weighting in panel A. Panel C is the finding — a small number of counties sit at the bottom of very large contributing networks, and those are precisely the places a co-location map cannot see.

How far downstream the data actually lets us follow. Three limits are worth stating plainly. (i) The closure is topological, not hydraulic: it says water flows A→B, with no travel time, discharge, dilution or decay. (ii) It is unbounded — there is no distance cut-off in the graph, which is why every claim in this report carries a tier label. (iii) It is geographically incomplete: the network is 507,486 reaches across 348 HUC8 subbasins in 10 HUC2 regions, concentrated in the Upper Mississippi, Ohio, Great Lakes, Tennessee and Northeast basins. Flood cells in the Gulf, South Atlantic, Texas and Pacific regions have no routable network in the federation and appear in this study only as co-location, never as a source. Every absence in the Southwest and the Southeast is a data-coverage absence, not a finding.

5.3 The typology: who keeps their risk, who receives someone else's

Combining the two axes — co-located flood-exposed facilities (retained) and tier-weighted upstream flood-exposed facilities (imported) — over the 519 counties that appear on either axis gives four groups.

The geography is legible at a glance: Imported and Compound counties trace the Ohio River from Cincinnati to its mouth, the Upper Mississippi from the Twin Cities south, and the Minnesota–Wisconsin reach, while Retained counties are scattered metros. This is the signature of a directed process, not of a spatial autocorrelation artefact.

Interactive map. Flood-cascade typology, OpenStreetMap-tiled. Every county is a clickable marker; the popup gives its typology, co-located flood-exposed facility count, tier-weighted imported upstream facility count, number of contributing upstream counties, RUCC, population, downstream monitoring coverage, and the federation KGs that contributed. Coordinates are the median position of the county's S2 Level-13 cells in spatialkg; basemap © OpenStreetMap contributors.

5.4 Who lives downstream

The imported group is not a random subset of American counties. It is more rural, smaller, and poorer-served than the retained group by every measure available in the federation.

Imported counties are 62% rural against 33% for Retained; their median RUCC is 6 against 3; their median population is 20,813 against 107,215 — roughly a fifth the size. In plain terms: the places that generate flood-mobilised contamination are metropolitan, and a large share of the places that receive it are small and rural. That asymmetry is the substantive result of adding the routing step, and it is invisible to any within-boundary burden metric.

5.5 How much the ranking changes when you drop the routing

The counterfactual the study was built to answer. Ranking the same 519 counties on co-located flood-exposed facilities alone — the co-location baseline — and comparing to the routing-aware cascade rank:

• 31 of the top 50 counties change (19 of 50 survive in both).
• Half of the top 100 change (50/100 overlap).
• Spearman ρ = 0.578, Kendall τ = 0.475 — a substantial but far from complete

correspondence.

• **179 counties (18.5 million residents) score exactly

zero on the baseline and non-zero with routing.** For these places the baseline is not approximately wrong; it is structurally blind.

The largest climbers are exactly the places the framing predicted: Lincoln County MO (+174 places, 1 co-located facility, 3,410 upstream), Wabasha County MN (+173), Dakota County MN (+173, zero co-located facilities and 1,885 upstream), Livingston and Ballard Counties KY (+172/+171, zero co-located, 4,742 upstream from 62 distinct upstream counties). Several are RUCC 8–9 — among the most rural classifications in the scheme.

6. Domain analyses

Declared coverage of the source families. The brief named four contaminant-source families. We ran three at cell resolution — regulated industrial facilities (§5.1), PFAS/contaminant observations (§6.3), water wells as a direct pathway (§6.2) — and could not run the fourth, manufacturers, at any useful resolution (§6.4). Agricultural land is treated in §6.4 as well. Nothing in this list is silently omitted.

6.1 The quadrant view

The mass along the left edge (retained score ≈ 0, imported score high) is the Imported group and is the visual statement of the whole study: a large, well-populated band of counties whose entire flood-contamination exposure comes from somewhere else. The Compound group in the upper right is small (59) but contains the highest-stakes places — Madison County IL, Jefferson County KY, Ramsey County MN, Hamilton County OH — which are both large emitters and large receivers.

6.2 Flooded drinking-water wells — a direct pathway, not a routed one

Wells deserve separate treatment because the pathway is direct: a flooded wellhead is a contamination route with no transport step to weight or attenuate. hydrologykg places 1,006 water wells inside modelled flood cells — 982 Illinois State Geological Survey wells and 24 Maine Geological Survey wells. Their recorded purposes matter:

• 333 Illinois wells with purpose WATER (water supply) and **23

Maine wells with use Domestic** — 333 + 23 = 356 wells whose function is to supply drinking water, sitting inside a modelled flood footprint.

• 419 monitoring wells and 211 engineering wells, plus 13 water-test, 3 irrigation

and a handful of stratigraphic/dry holes.

This count is a floor, and a low one: hydrologykg carries well inventories for only two states, so 356 flood-exposed drinking-water wells is what two states' worth of coverage yields, not a national estimate. Read it as a demonstration that the pathway is queryable and materially populated, not as a national figure.

6.3 Downstream monitoring — where the receiving end is unwatched

Of the 25,063 downstream S2 cells reached by the routing, only 727 contain a sawgraph contaminant-observation or monitoring feature, spread across 79 counties. Turned around: 151 of the 208 Imported and Compound counties — 8.4 million residents — have no contaminant monitoring anywhere in their downstream cells in this federation.

The mismatch is even starker at the source end. sawgraph carries 88,007 S2 cells with contaminant features, but only 13 of them coincide with a UF-OKN flood cell. PFAS observation coverage and modelled flood exposure are, in this release of the federation, almost disjoint: sawgraph is dense in Maine and in scattered WQP sites in Indiana, Minnesota, Arizona and Illinois, while the flood footprint is dense in Kentucky, Florida, Michigan and Wisconsin. That is a monitoring-design finding in its own right — the places we model as flooding are not the places we sample for PFAS — but it also means this study cannot corroborate any routed link with a measurement. Every downstream claim here is topological.

6.4 Source families we could not place: manufacturers and agricultural land

Manufacturers (sudokn). The federation's verified crosswalk documents a computed S2 bridge for sudokn covering ~42,560 sites. In the release queried here, only 225 sites carry hasLatitudeValue/hasLongitudeValue, and those that do are overwhelmingly foreign semiconductor headquarters rather than US small and medium manufacturers. We therefore could not place sudokn manufacturers in the flood footprint at cell resolution and have excluded them rather than substitute a state-level proxy that would not be comparable to the cell-resolution facility layer. EPA FRS manufacturing (NAICS 31–33, 799 facilities, §5.1) covers the regulated subset of the same population.

Agricultural land (sockg). SOC-KG is a soil-carbon research-site graph, not a land-cover layer: its verified spatial footprint is ~1,069 S2 cells across 62 counties nationally. That is two orders of magnitude too sparse to characterise agricultural land in a 2,738-cell flood footprint, and we deliberately did not run a cell-level agricultural intersection on it. The agricultural signal that is present in this study comes from fiokg — 88 flood-exposed facilities in NAICS 11 (agriculture, forestry, fishing and hunting) and the EPA animal-operations and pesticide programme interests.

7. Discussion

Three things follow from the analysis, in descending order of confidence.

First, and most securely: the routing step is not a refinement, it is a different map. Thirty-one of the top fifty counties change when it is added, and 179 counties move from a structural zero to a positive score. Any burden index built on within-boundary co-location will systematically under-serve the downstream half of a watershed. This claim rests only on graph topology and is as strong as the NHDPlus network itself.

Second, with good confidence: imported flood-contamination risk is a rural burden. The Imported group is 62% rural against 33% for Retained, with a median population a fifth the size. The mechanism is not mysterious — regulated industry clusters in metros, metros sit on rivers, and rivers run to smaller places — but the magnitude is worth naming, and it inverts the usual urban framing of industrial environmental burden. The caveat is that ruralkg's RUCC and population are a 2013 vintage.

Third, as a hypothesis to test rather than a conclusion: the receiving end is systematically unmonitored. 151 Imported/Compound counties have no downstream contaminant monitoring feature in this federation. Some of that is real monitoring absence and some is federation coverage — sawgraph is a PFAS-focused graph with two dense states — and the two cannot be separated with the data here. It is nonetheless the sharpest actionable signal the study produces.

What this supports operationally. A prioritisation for post-flood sampling would look different from a prioritisation for site hardening. Hardening should target the Compound and Retained counties — Madison County IL, Jefferson County KY, Ramsey County MN, Hamilton County OH — where large numbers of flood-exposed facilities sit. Sampling should target the Tier-A Imported counties with zero downstream monitoring: Wabasha and Dakota Counties MN, Lincoln County MO, Livingston and Ballard Counties KY. And the 356 flood-exposed drinking-water wells (§6.2) are a discrete, enumerable population that could be tested directly.

Testable predictions. (i) Post-flood surface-water sampling in Tier-A Imported counties should detect elevated metals, chlorinated solvents and PFAS relative to matched Tier-C counties. (ii) The electroplating, automotive-repair and dry-cleaning facilities identified in §5.1 should dominate any source apportionment of a flood-mobilised solvent/metals signal in these corridors. (iii) Flooded private wells in the Illinois WATER-purpose set should show microbial and, near plating and fuel-handling premises, chemical exceedances after inundation.

8. Comparison with prior work

According to PubMed, retrieved via the PubMed MCP connector, the claims below were checked against the primary literature on flood-induced contaminant mobilisation, flood-exposed hazardous-waste sites, and post-flood private-well contamination. The full per-claim record with citations is in Flood-Cascade_literature_comparison.md.

Claims 1–6 were checked against the full abstracts returned by PubMed; none of the eleven claims was verified against article full text, so no reference in §12 carries a full-text-verified marker.

Where the KG evidence diverges from the literature. The divergences are of scope, not of fact. The literature establishes the mechanism — flood waters mobilise and redeposit contamination (Claim 2), flooded wells become contaminated (Claim 4), and legacy-contaminated environmental-justice communities receive redistributed pollutants (Claim 3) — but works at the scale of a single event, a single basin, or a single community. This study contributes the complementary scale: a national, reproducible, source-to-receptor accounting over a hydrologic network. Two divergences are worth flagging as graph rather than literature problems. First, sudokn's coordinate coverage in this release (§6.4) contradicts the federation's own verified crosswalk figure of ~42,560 placed sites — that is a data regression to report upstream, not a finding. Second, the near-disjointness of sawgraph and ufokn coverage (Claim 11) is a property of which states each project has loaded, and should not be read as a statement about US PFAS monitoring.

9. Full ranked results

The complete ranked table — all 519 counties with both risk axes, the typology, the tier label, rurality, population, downstream monitoring coverage and both rankings — is in Flood-Cascade_results.xlsx (sheet Ranked Results), alongside sheets for the industry breakdown, the EPA programme inventory, the flooded-well register, a sample of the routing links, and a Methods & Rules sheet. Intermediate extracts are in data/.

Tip: click any header to sort; use the search box for a county or state; use the pull-downs to isolate a typology, a hydrologic tier, a state, or the rural flag. The sources (n) column counts how many federation KGs contributed to that row — ufokn supplies the flood footprint, fiokg the co-located facilities, hydrologykg the routing, spatialkg the county geography, ruralkg the rurality and population, and sawgraph the downstream monitoring.

A representative slice — the ten counties with the highest imported-risk load and no co-located flood-exposed facility at all:

Two patterns stand out. The Kentucky/Illinois cluster at the top is a Tier B mainstem signal — very large upstream loads, many contributing counties, but attenuated by distance down the Ohio; the routing puts them high, the tier label says treat with care. The Minnesota pair is the opposite and the more actionable: Tier A, same-subbasin connections with fewer upstream counties, which is a much stronger claim about a much shorter transport path — and Dakota County MN carries 398,552 residents with zero flood-exposed facilities of its own.

10. Summary of findings & limitations

Findings recap. Across a national modelled flood footprint of 47,512 buildings (2,738 S2 cells), 11,085 EPA-regulated facilities sit inside the water's way, with exposure moderately concentrated in manufacturing and small chemically-active service premises (HHI 908.0; top-3 sectors 45.1%). Routing those sources through the federation's NHDPlus downstream closure produces 996,529 links to 15,727 downstream reaches and a four-way county typology: 153 Retained, 149 Imported, 59 Compound, 158 Low. The imported group is the finding — 179 counties and 18.5 million people carry flood-contamination exposure generated entirely upstream, they are 62% rural against 33% for the retained group, and 151 of the Imported/Compound counties have no downstream contaminant monitoring at all. Dropping the routing step changes 31 of the top 50 counties. Separately, 1,006 water wells — 356 of them drinking-water supply — sit inside the modelled flood footprint as a direct, unrouted pathway.

Limitations.

1. The flood layer is model output, not observation. UF-OKN publishes forecast/scenario

flood-depth predictions for buildings; a "flood cell" is a cell where the model predicts building inundation, not a place that has flooded. Nothing here is a record of an actual flood.

1. The routing is topological, not hydraulic. downstreamFlowPathTC asserts connectivity with no

travel time, discharge, dilution, sorption or decay. The HUC-based tiers are a proxy for distance, not a transport model. A Tier A link is a better bet than a Tier C link; neither is a concentration estimate.

1. The routable network is geographically incomplete. 507,486 reaches across 348

HUC8 subbasins in 10 HUC2 regions. Flood cells in the Gulf, South Atlantic, Texas, Great Plains and Pacific regions have no routable network here, so those states appear only as co-location. Absence of imported risk in the Southeast and Southwest is a coverage artefact.

1. Only 20% of flood-exposed facilities carry a NAICS code (2,270 of 11,085).

The industry composition in §5.1 describes the coded subset and may not represent the rest; the EPA programme-interest counts, which are far more completely populated, are the more robust view.

1. No contaminant release is asserted or observed. A facility inside a flood cell is a facility

that may be inundated; whether it holds anything mobilisable, in what quantity, and whether it would be released, is entirely outside this data.

1. The flooded-well count is a two-state floor. hydrologykg carries well inventories only for

Illinois and Maine; 1,006 wells is what those two states yield, not a national figure.

1. Monitoring coverage conflates real gaps with federation gaps. sawgraph is a PFAS-focused

graph dense in Maine with scattered national WQP sites; a county with "no downstream monitoring" may be monitored by programmes this federation does not carry.

1. Manufacturers and agricultural land are not represented at cell resolution (§6.4) — sudokn

coordinates and sockg extent are both far too sparse in these releases.

1. Rurality and population are 2013 vintage (ruralkg RUCC), so the social profile in §5.4 lags

the flood and facility layers by more than a decade.

1. The typology cut (0.60 percentile on each axis) is a analyst choice, not a natural break. The

continuous scores are in the workbook; group sizes shift with the cut, though the qualitative rural/urban asymmetry is stable across reasonable alternatives.

1. Both risk axes count facilities, not hazard. A dry-cleaner and a chemical plant count one

each. A hazard-weighted version would need release inventories this study did not join.

1. County centroids in the maps are the median position of a county's S2 cells, not a

population-weighted or geometric centroid, because the federation carries no county polygons; use them as locators, not as geometry.

11. Reproducibility

Everything needed to replicate this analysis — the originating prompt verbatim, the replicator specification (selection rules, thresholds, join recipes, tier weights, verified quantities and limitations), every supporting SPARQL query with its row count, and the pinned KG versions and timing — is in Flood-Cascade_reproducibility.md, with the exact scripts in scripts/ and the intermediate extracts in data/.

12. References

Retrieved via the PubMed MCP connector. Full-text verification via the Paperclip MCP connector.

1. Amolegbe SM, et al. Adapting to Climate Change: Leveraging Systems-Focused Multidisciplinary Research to Promote Resilience. International journal of environmental research and public health. 2022. PMID:36429393 · doi:10.3390/ijerph192214674
2. Singer MB, et al. Enduring legacy of a toxic fan via episodic redistribution of California gold mining debris. Proceedings of the National Academy of Sciences of the United States of America. 2013. PMID:24167273 · doi:10.1073/pnas.1302295110
3. Chukwuonye GN, et al. Source attribution of polycyclic aromatic hydrocarbons and dioxins in soil and dust following compound climate events in legacy-contaminated environmental justice areas. The Science of the total environment. 2026. PMID:42456624 · doi:10.1016/j.scitotenv.2026.182028
4. Pieper KJ, et al. Microbial Contamination of Drinking Water Supplied by Private Wells after Hurricane Harvey. Environmental science & technology. 2021. PMID:34032415 · doi:10.1021/acs.est.0c07869
5. Mapili K, et al. Occurrence of opportunistic pathogens in private wells after major flooding events: A four state molecular survey. The Science of the total environment. 2022. PMID:35182640 · doi:10.1016/j.scitotenv.2022.153901
6. Peer K, et al. The private well water climate impact index: Characterization of community-level climate-related hazards and vulnerability in the continental United States. The Science of the total environment. 2024. PMID:39510280 · doi:10.1016/j.scitotenv.2024.177409
7. Proto-OKN federated SPARQL endpoint (FRINK), accessed 2026-07-26 via the mcp-okn server. KG versions pinned in §2.