Purpose
Why This Exists
This is a citizen science project. We built it because the people who live in the path of western wildfires and droughts deserve access to the same data and analysis that federal agencies and insurance companies use to make decisions about their lives — for free, without a paywall, without a login, without asking permission.
What we commit to
- Data accuracy above everything. Every number traces to a specific federal, state, or provincial data source. When we make a mistake, we correct it publicly. When a claim is uncertain, we say so. We will never overstate risk to generate attention or understate it to avoid controversy.
- Boots on the ground.This is not a dashboard built from a desk. We verify conditions in the field, cross-reference agency reports with local observations, and update analysis based on what's actually happening — not just what models predict.
- Real policy analysis. We read the bills, track the budgets, and grade the agencies on whether their actions match the threats the data identifies. Policy grades are editorial assessments, clearly labeled as such.
- A self-healing data network. Federal data sources are being defunded, restructured, and shut down. When the government turns off a data source, we find alternatives. When those get turned off, we derive equivalent insights from other sources — satellite imagery, state agencies, academic repositories, allied nations. The data will keep flowing regardless of who controls the servers.
- Always free, always honest. No subscriptions. No premium tier. No sponsored content. No softening findings to protect relationships. The people in fire and drought country are the audience. Everyone else is welcome to read along.
Firewatch Engine
Our prediction model — Firewatch — is an ensemble machine learning system trained on 37 million observations across the western fire corridor (WA, OR, ID, MT, CA, CO, WY, UT, NV, NM, AZ + BC). It learns from 47 years of daily climate, 1,830 SNOTEL stations across 11 states, 278,000 fires with cause classification, fuel maps, soil moisture, snow disappearance timing, and topography to identify where large fires are most likely to burn next.
91% accuracy means: when Firewatch ranks an area as high-risk, it's correct 9 out of 10 times compared to areas it ranks as low-risk (technical metric: 0.911 AUC-ROC on a held-out test set the model never saw during training). It predicts where large fires (>1,000 acres) concentrate — not individual ignitions. 2026 is the live public validation period.
Top predictors (SHAP importance, v2.1)
15. Snow disappearance date (novel v2.1 feature)
Data sources (v2.1)
| Dataset | Source | Coverage |
|---|---|---|
| Daily climate (VPD, fuel moisture, temp, precip, wind, burning index) | gridMET (U of Idaho) | 2000–2025, 4km, 32–49°N |
| Wind direction (east wind frequency) | gridMET (U of Idaho), 4km daily | 2000–2025, % of summer days with easterly flow |
| Snowpack (daily SWE) | NRCS SNOTEL + SNODAS | 1,830 stations, 11 states, 1981–2026 |
| Snow disappearance date | Computed from daily SNOTEL | 2000–2025 (novel feature) |
| Soil moisture (8-inch depth) | SNOTEL SMS sensors | 55 stations, summer 2000–2025 |
| Fire perimeters | NIFC WFIGS + MTBS | 6,853 perimeters, 11 states, 1992–2024 |
| Burn severity | USGS MTBS | 7,949 fires >1,000 ac, 1984–2024 |
| Fire cause | FPA-FOD | 312,000 fires, lightning vs human |
| Canadian fires | NFDB | 155,198 BC records + 1,878 polygons |
| Fuel type + canopy | LANDFIRE | FBFM40, EVT, canopy cover, 30m |
| Topography | GMTED2010 DEM | Elevation, slope, aspect, 250m |
| Climate indices | NOAA CPC | ENSO Nino 3.4, PDO, ENSO velocity |
| Reservoir storage | USBR Hydromet, USACE, CDEC | 39 reservoirs, 11 states, real-time |
| Air quality (PM2.5) | EPA AQS + PurpleAir | 700 monitors + 11,500 sensors |
| Summer temperature | NOAA NCEI | 131 years (1895–2025), 11 states |
| Smoke feedback | EPA AQS annual PM2.5 | Prior-year smoke as predictor (novel) |
Model history
| Version | AUC | Key change |
|---|---|---|
| v0.1–0.3 | 0.67–0.72 | PNW annual grid, added topography + LANDFIRE + fire cause |
| v1.0 | 0.841 | Monthly resolution (biggest single improvement) |
| v1.1–1.2 | 0.90–0.91 | Full western corridor, lag features, interaction terms, ensemble |
| v1.3 | 0.913* | SNODAS gridded SWE. *Leaked metric — test set used during tuning. |
| v1.4 | 0.901 | Proper 3-way split, ENSO/PDO, held-out test. |
| v2.0 | 0.909 | 11-state SNOTEL, SWE anomaly, VPD×SWE interaction. |
| v2.1 | 0.911 | Snow disappearance, east wind, prior-year precip, reservoir anomaly, ENSO velocity, soil moisture, smoke feedback, cheatgrass×precip. 50 features. Production model. |
Compared to published models
Station-level, no spatial prediction
25km global, monthly (GMD). Uses r not AUC-ROC — 0.80 is approx.
Monte Carlo simulation
Daily, 375m, not public
4km monthly, 50 features, held-out test, snow disappearance, soil moisture, ENSO velocity, smoke feedback, public data only
Known limitations
- Predicts large fires (>1,000 acres) only — not ignitions or small fires.
- Wind at 4km daily average — extreme events (2020 Labor Day) underrepresented.
- SWE methodological discontinuity at 2010 (SNOTEL IDW → SNODAS gridded).
- BC predictions use nearest-grid climate interpolation.
- Smoke forecast is distance-based, not fuel-composition-aware. Smoke production varies 5x depending on what burns (grass vs old-growth vs organic soil). A proper resolution smoke model requires mapped fuel condition, plant stress from VPD, live fuel moisture (NDVI), and sub-4km atmospheric mixing height — all on the v3.0 roadmap.
- This is a research platform, not an operational tactical tool for evacuation or life-safety decisions.
Open source
All code, data pipelines, and trained models are available for inspection and replication. No proprietary data. No black boxes. Every number on this site traces to a specific public data source. If you find an error, tell us. We'll fix it the same day.