The Irrigated Desert

Base map Data opacity 80%

Time

Click any hexagon for its detail. This map describes site conditions — standing water, containers, year-round irrigation — not the people who live there.

Pick a spot

Choose your ZIP code, tap Use my location, or click the map. You will get a separate reading for the day-biting Aedes aegypti and for the evening Culex mosquitoes that carry West Nile virus.

Season timeline

When they are out

The bad years, and what actually made them bad

Between 2008 and 2025 the valley had three West Nile seasons far outside the rest: 2015, 2019 and 2024. Five seasons were effectively silent — 2010, 2016, 2018, 2021 and 2022. Years are classified from the data, not chosen by hand: an outbreak year is one whose season-long pool positivity sits more than one standard deviation above the mean, a quiet year is one under 0.5%.

It was not more mosquitoes

This is the finding that surprises people, and it is the strongest one here. Pooled across years, outbreak seasons produced 16.6 mosquitoes per trap against 15.7 in silent seasons — a six per cent difference — while infection differed 135-fold. Across all eighteen years, mosquito abundance carries essentially no relationship to how much virus circulated.

It started earlier, and spread wider

Year by year

Year		Positivity	Per trap	First positive	Hexes with virus	Pools tested

What does not explain it

Ten annual climate summaries from the project's own 800 m record were tested against season positivity. None survives. The strongest is —, and its sign is backwards — more transmission-optimal days going with less infection — which at seventeen years and ten comparisons is what noise looks like. Whatever separates a 2024 from a 2022, an annual climate average does not capture it.

Annual climate variable	r vs positivity	p

Where this lands

How Las Vegas compares

Aridity is not protection. Las Vegas receives about 93 mm of rain a year, yet supports both an established West Nile cycle and, since 2017, breeding Aedes aegypti — because in a desert the water that matters is the water people supply. The cities below test how far that pattern travels. Each is measured the same way, from the same source, over the same 41 years: one airport weather station per city, NOAA Global Summary of the Day, 1984–2024, using the identical mosquito thresholds applied inside the valley. The density maps are all 60 km squares from the same two global layers on the same shared scales — Sentinel-2 imagery at 80 m and population density at 1 km. Click any city to open it larger, with zoom and pan.

Compare on Map layer

Two cities carry a caveat

Yuma has no single weather station spanning the period. Its airport's record is split across renumbered station IDs, so the series here is spliced from two (1984–1999 and 2005–2025, with a gap around 2000–2004). Its levels are sound and sit exactly where a hot desert city should — between Phoenix and Palm Springs on every heat measure. Its trend is not: splicing instruments and siting produces −3.1 Culex-nights per decade against +1.8 to +10.9 for its neighbours. That is an artefact of the join, not a cooling desert, and the same anomaly was flagged in this project's earlier station work. Yuma is therefore shown for levels and withheld from every trend comparison.

Dhaka is not fragmented, just late. Its station reports about six usable days a year before 2000 and 272–366 after, so its record is 2001–2024, twenty-one years rather than the forty-one every other city gets. Means are comparable; the trend is fitted over its own shorter window and is flagged rather than pooled with the rest. Its rainfall is shown as absent, not as zero: the station logs precipitation observations on 239–268 days a year and records 0 mm on every one of them, which for a city receiving roughly 2,000 mm is a reporting gap, not a dry climate.

Cells marked no comparable data are genuinely empty, not zero. West Nile virus is under systematic surveillance in the United States and the European Union and nowhere else in this panel; reporting an absent surveillance system as zero cases would invert the comparison.

Water, aridity, and what this cannot tell you

This dashboard sits on top of a research project on mosquito ecology in the Las Vegas valley, run with Louisa Messenger's laboratory at UNLV. Two grids cover the valley: 18,125 hexagons of 250 m for Aedes aegypti, and 1,184 hexagons of 1 km for the Culex species. Onto them are joined 31 years of 800 m daily climate, 102,567 swimming pools mapped from 0.6 m aerial imagery, 664,728 building footprints, land cover, irrigation greenness, flood-control infrastructure, and two decades of mosquito trapping and West Nile testing by the Southern Nevada Health District.

The most important limit

In spatially blocked cross-validation, habitat variables did not improve on space and season alone — −0.022 for Aedes, −0.001 for Cx. quinquefasciatus, +0.028 for Cx. tarsalis. These maps can tell you where mosquitoes are likely to be. They cannot yet tell you why. Treat every surface here as an operational targeting layer, never as a mechanism.

Why the Culex map looks flat

Because it is. Modelled Cx. quinquefasciatus suitability spans only 0.729–0.786 across the entire valley, and Cx. tarsalis only 0.238–0.336. That near-uniformity is the finding rather than a failure: for the main West Nile vector, essentially the whole developed valley is suitable. So the Culex bands on this map are not built from modelled habitat at all — they come from what was actually trapped and tested since 2015. Where no traps run, the map says so instead of guessing.

Exploring by time — and why it is a different quantity

The default map is model output. The moment you pick a time period it becomes observed trapping — what the Southern Nevada Health District actually caught and tested in those weeks. The two are not interchangeable, so the map relabels itself and the "which leads here" layer switches off, because dominance is a comparison of two modelled surfaces and has no observed equivalent across two different grids.

Every abundance figure in the time view is per trap record. This is not a stylistic choice: raw weekly counts correlate 0.648 with the number of traps run, so an untreated "hot" week is often just a well-trapped one. Dividing by trap records drops that correlation to 0.152.

Resolution is limited by what was sampled, not by the interface. There are 10,513 Culex hexagon-week-year cells across 2008–2025, but the median hexagon-week carries a single observation, so a hexagon is banded only where at least three trap records stand behind it and everything else is drawn as unsampled. Coverage also grew enormously: 74 hexagons were trapped in 2008 against 506 in 2024, which is why a single early year looks emptier than a recent one. Aedes weekly trapping only becomes substantial from 2022. Narrow slices — one species, one year, one week — are honest but often too thin to read as a map, and the banner says so when that happens.

One pattern falls straight out of the timeline and is worth stating on its own: Culex abundance peaks around epidemiological week 22, but the share of mosquito pools carrying West Nile virus peaks around week 26. The most mosquitoes and the most virus are not the same week, so a control programme timed to peak abundance is timed slightly early for peak transmission.

Where the map goes quiet

Hexagons more than 2 km from any trap are masked. That is — of the Aedes grid, and — of the Culex grid has too little recent trapping to band. Those areas are drawn in hatched grey and labelled not enough local trapping. They are not low-risk areas. They are unmeasured ones, and the distinction matters more than any number on this page.

Water, wealth, and what actually drives this

The valley's mosquito problem runs opposite to the usual public-health pattern. In a desert city, the water that breeds mosquitoes is water people supply. Head to head, the measure that survives is pool area; applied irrigation does not. Household income predicts pool area, which is why income appears in the models at all — and the effect of pools is larger in wealthier neighbourhoods.

Read that carefully, because it is easy to read wrong. The models respond to standing water at container scale, not to the people near it. A pool, a neglected fountain, a saucer under a plant and a tarp holding rain are the same object to a mosquito. Nothing in this work identifies any community as a source of mosquitoes or of human cases, and this dashboard is built so it cannot be read that way: the resident map carries physical site conditions only — no income, no ethnicity, no case counts.

Demographic data has one legitimate use here, and it is the opposite of blame: identifying who carries the burden so that spraying, source reduction and outreach go where illness actually lands. It is not a predictor in any model on this page. Estimated human cases are aggregated to ZIP code and never shown at hexagon level, both to protect privacy and to keep any single neighbourhood from being read as a culprit.

What the seasonality curves actually measure

The "when they are out" chart is not a share of total mosquito activity, and the three curves are not standardised against each other. Each comes from its own binomial model fitted to that species alone — present ~ s(doy, cyclic) + s(x, y) — predicted at the valley's median trap location. The quantity is the modelled probability that a trap at a typical site detects that species on a given day of year. The underlying file stores it as a log-odds; the page converts it to a percentage because log-odds are not readable.

That makes exactly one comparison safe and one comparison qualified. All three curves share the same units, so they can sit on one axis — but detection also depends on trap type, and trap types are targeted at different species. So timing is directly comparable; height is only indicative. The chart defaults to scaling each species against its own peak for that reason, with a toggle to absolute detection rates for anyone who wants them. At peak, the model puts detection at about 66% for Ae. aegypti, 77% for Cx. quinquefasciatus and 28% for Cx. tarsalis; the floors are 0.6%, 16% and 6% respectively, and that quinquefasciatus floor is what "close to year-round" means.

Note also that this is presence in a trap, not abundance. It answers "is this species around this week", not "how many". The season timeline above answers the second question, from raw counts per trap record.

Three vectors, three different answers to rain

Temperature acts immediately — Aedes trap catch rises +0.122 per °C at lag 0. Rainfall acts two to four weeks later, through recruitment. The three species do not agree with each other about rain, which is why they are modelled separately here rather than pooled into one "mosquito risk" number.

The global comparison, and its honest gaps

Las Vegas is warming at night faster than almost anywhere in its own region, and the cities it is converging on already live with Aedes-borne disease. That is the argument the comparison tab makes. It holds only for layers that are genuinely comparable:

Layer	Comparable?	Source and caveat

Two layers deserve a warning in bold. GBIF occurrence records measure who went looking, not how many mosquitoes there are — Las Vegas returns thousands of Culex records because it runs an intensive surveillance programme, while Mumbai returns a handful. Read that column as sampling effort. And ethnicity has no globally comparable construct, so it appears nowhere in this panel; substituting a proxy would hollow out the very argument the comparison exists to make.

The base maps

A published page here runs under a strict content policy that blocks every external host, so no tile server — Esri, Google, Mapbox, OpenStreetMap — will load. The base maps are therefore embedded, which restricts them to imagery that may be redistributed. Both sources are US federal public domain: satellite is NAIP 2022 at 0.6 m from the USDA, resampled to 15 m for the valley — the same imagery this project already used to map 102,567 swimming pools, so it is the study's own evidence rather than a decorative backdrop. Streets, boundaries and water are Census TIGER/Line 2023, shipped as vectors rather than a rendered image so they stay sharp at every zoom level.

The arterial grid is selected by total length per named street, not per segment: TIGER splits roads into pieces with a median length of 157 m, so filtering segments individually discards every major road in the valley. 340 named streets survive that test, out of 63,716 road features.

The city panels

Each comparison city is drawn as the same 60 km square, from the same two global layers, on the same two shared scales: Sentinel-2 L2A imagery at 80 m (Copernicus/ESA, open licence) and GHSL population density at 1 km, the latter on a shared log scale so a darker city really is denser. Clicking a city opens it larger with zoom and pan. Political boundaries and coastline come from Natural Earth; at 60 km most cities have none, but the two that matter here do — San Diego and El Paso each sit in a single urban centre split by an international border, which is why the urban-centre database files them under Tijuana and Ciudad Juárez.

Resolution is set by the page budget, not by preference: this account has no out-of-page asset store, so imagery must be embedded in the file itself. Measured, 30 m imagery would have cost 25 MB across nineteen cities and 60 m about 6.6 MB; 80 m costs roughly 3 MB and is still 12.5 times finer than the 1 km density grid these panels used before. Scene selection ranks on cloud and missing data together — ranking on cloud alone actively prefers near-empty scenes, since a scene covering a sliver of its tile is trivially cloud-free, and that error had Cairo at 22% coverage.

Weather stations behind the comparison

Station choice is automatic, not hand-picked, because hand-picking already cost this project once. Years with fewer than 328 observation days are dropped and logged rather than averaged, since count metrics scale with days observed and a ragged record manufactures false trends. Las Vegas active nights come out at — per decade here, against +8.10 per decade from independent 800 m gridded climate over the shorter 1994–2024 window.

City	Station	ICAO	km	Years	Dropped

Airport stations sit in hot, dry, low-albedo settings and read warmer at night than residential neighbourhoods. Use them to compare cities with each other, not as absolute neighbourhood temperatures. Alice Springs falls below the population threshold of the global urban-centre database, so its city statistics are blank while its climate record is complete.