HeatShield 26 scores 3,434 H3 resolution-9 hexagons (0.1053 km² each) covering Downtown, Midtown, Montrose, River Oaks, Uptown/Galleria, Rice/West University, the Texas Medical Center, NRG Park, East End, Near Northside, the Heights and the University of Houston — inside Loop 610 and a short buffer beyond it. The Heat Event Risk Index is a decision-support composite, not a medical prediction. It ranks where the same heat is most likely to harm the most people, so planners can compare options consistently across any of 6 real Houston venues.
Heat Event Risk = ( 0.40 × Heat exposure
+ 0.25 × Event crowd exposure
+ 0.20 × Health & social vulnerability
+ 0.10 × Shade deficit
+ 0.05 × Cooling access deficit )
× (1 − protective factor from water/misting stations)Every component is normalized to 0–100. Weights are adjustable in the planner and re-normalized to sum to 1. Tiers:
| Component | How it is computed |
|---|---|
| Heat exposure | Scenario heat index from forecast air temperature and humidity (NWS Rothfusz regression) plus the cell’s H3AT heat-index anomaly for the chosen event window. Scaled linearly from 80°F (NWS “Caution”) = 0 to 125°F (“Extreme Danger”) = 100. |
| Event crowd exposure | Modeled event person-hours in the cell (see crowd model). Log-scaled: 100 × log₁₀(1 + person-hours) / log₁₀(1 + 5,000), capped at 100. |
| Health & social vulnerability | 50% CDC SVI 2022 overall percentile (Texas) + 50% health-burden percentile: the mean study-area percentile of coronary heart disease, COPD, diabetes, chronic kidney disease (CDC PLACES), age 65+ and disability (ACS via SVI). Tract values are assigned to cells by centroid. |
| Shade deficit | 100 × (1 − canopy ÷ 40%), floored at 0. Canopy from USFS NLCD Tree Canopy Cover (30 m), which under-counts street trees relative to LiDAR but covers every cell consistently; temporary shade adds canopy-equivalent points. |
| Cooling access deficit | Straight-line distance to the nearest City of Houston cool center (or a placed cooling hub), scaled so 1 mile or more = 100. |
Heat anomalies come from H3AT’s modeled heat-index rasters, sampled at 7 points per cell. The 2020 layers cover 2,625 of 3,434 cells. The published 2024 layers cover disadvantaged-community focus areas, and where both exist the two campaigns disagree at neighborhood scale (afternoon r = -0.03 across 1,502 shared cells). So HeatShield does not blend them. H3AT 2020 is the primary surface, measured as °F above or below its study-area median. 2024 values fill only the cells 2020 misses, after removing the mean 2024−2020 offset.
For cells neither campaign covers, inverse-distance interpolation (3 km radius, shrunk toward the median) is tested against a flat median fill with spatial hold-out cross-validation: every measured cell is predicted from measured cells at least 989 m away, the typical distance from a gap to data. Interpolation is used only where it wins. A Landsat surface-temperature regression was also tested and explains little of the afternoon anomaly.
| Window | 2020 cells | 2024-only | Filled | 2020↔2024 r | CV RMSE interp. / median | LST reg. R² | Gap fill used |
|---|---|---|---|---|---|---|---|
| Morning | 2,625 | 266 | 543 | 0.47 | 2.05°F / 2.55°F | 0.53 | Interpolation |
| Afternoon | 2,625 | 266 | 543 | -0.03 | 0.95°F / 1.02°F | 0.01 | Interpolation |
| Evening | 2,625 | 266 | 543 | 0.22 | 0.99°F / 1.06°F | 0.13 | Interpolation |
Takeaway: afternoon heat-index differences inside this corridor are modest (most cells fall within about ±1°F of the median), so heat mainly sets how dangerous the whole event day is. Crowd exposure, vulnerability, shade and cooling access decide where the danger concentrates. Each cell’s heat source is labeled in the planner.
No public pedestrian counts exist for any of these events, so HeatShield routes an attendance scenario instead of pretending to know. The algorithm is identical for all 6 venues — only the geometry (gates, nearby lots, nearest rail stations) differs, computed fresh from OpenStreetMap and METRO data rather than hand-typed per venue. Scenario defaults (NRG Stadium): 68,000 attendees, arrival split 35% rail, 40% car, 15% rideshare, 10% walk; 50% of car arrivals tailgate for 2 h. All are sliders in the planner, per venue.
Routes are straight lines, not a street network — adequate inside a venue’s own lots, coarser across downtown blocks. “Visitors routed through critical zones” counts attendees whose modeled route touches at least one critical cell.
What the pipeline computed per venue
| Venue | League | Capacity | Parking lots | Spaces | Nearest stations used (weight) |
|---|---|---|---|---|---|
| NRG Stadium | FIFA World Cup 2026 · NFL (Texans) | 68,777 | 44 | 53,381 | Stadium Park/Astrodome (57%), Smith Lands (22%), Fannin South (21%) |
| Daikin Park | MLB (Astros) | 41,168 | 87 | 13,997 | Convention District (60%), EaDo / Stadium (24%), Central Station / Capitol (17%) |
| Toyota Center | NBA / NHL exhibitions (Rockets) | 18,500 | 75 | 13,639 | Bell (36%), Convention District (35%), Main Street Square (29%) |
| Shell Energy Stadium | MLS / NWSL (Dynamo FC / Dash) | 22,039 | 32 | 4,333 | EaDo / Stadium (78%), Convention District (17%), Central Station / Capitol (5%) |
| TDECU Stadium | NCAA football (UH Cougars) | 40,000 | 54 | 13,470 | TSU / UH Athletics District (79%), Elgin / Third Ward (13%), UH South / University Oaks (8%) |
| Rice Stadium | NCAA football (Rice Owls) | 47,000 | 35 | 9,580 | Dryden / TMC (40%), Memorial Hermann Hospital / Houston Zoo (35%), TMC Transit Center (26%) |
| Intervention | Cost | Modeled effect | Feasibility | Legacy |
|---|---|---|---|---|
| ❄ Cooling hub | $150,000 | New cooling site (recomputes distance-to-cooling everywhere); heat exposure −8 in cell, −4 adjacent; protective factor −8% risk in cell (on-site EMS triage). | 3/5 | 5/5 |
| ⛱ Temporary shade structures | $80,000 | +35 canopy-equivalent points in cell, +15 adjacent; heat exposure −8 in cell, −3 adjacent (less direct sun on people waiting or walking). | 4/5 | 4/5 |
| 💧 Water & misting stations | $35,000 | Protective factor: final risk −8% in cell, −4% adjacent. | 5/5 | 2/5 |
| 🚌 Shuttle redistribution | $100,000 | Walking person-hours (not queues or tailgating) −50% in cell, −30% adjacent, −10% two cells out. | 3/5 | 3/5 |
Costs and effect sizes are placeholders for comparison, not vendor quotes or measured outcomes — replace them with City procurement and EMS data before real decisions. The optimizer minimizes exposure burden = Σ (event person-hours + residents × 0.25 h assumed outdoor exposure) × risk ÷ 100, adding the intervention with the best burden reduction per dollar until the budget runs out.
The only generative-AI feature is “Explain this recommendation”. It sends the numbers already shown in the zone panel to Google Gemini (or Groq when Gemini is unavailable) and asks for a 3–4 sentence rationale that uses only those facts. Scores, routing and optimization are deterministic code. Without an API key the planner shows a template explanation instead.
Data pipeline: scripts/build-data.mjs. Dataset built September 15, 2026.