Hazards
Hazards we track
This page lists every hazard on the map. For each one it gives what the hazard shows, where it applies, and its status. A hazard is live, in progress, or planned.
live on the map now ·
in progress we are building it ·
planned not started
| Hazard | What it tells you | Where | Status |
|---|---|---|---|
| Urban heat | Daytime satellite surface temperature of each area, measured against the least built-up ground in the same map extent, and the reasons for the difference. In the pre-monsoon season dry open ground is often hotter at the surface than the built-up city, so many urban areas read below their reference. This is surface temperature, not the air temperature people feel. | 26 cities. The reference is unreliable in 5 of them, which the map marks. | Live |
| Flood | Which streets hold water after rain, from terrain and drainage data. Each point gives its reasons. The map also shows recorded flood points. | 4 cities, about 670 m resolution | Live |
| Dams and rivers | Large reservoirs and the river network. Click a dam to trace the water downstream. Three of six dams have a measured fill level from CWC gauge readings. The other three have no full-reservoir figure, so we cannot give a percentage. | North India: 5,418 river sections, 4,185 canals | Live |
| Ward boundaries | Ward and sector outlines. These are reference shapes. They are not land-use or zoning maps, which are not available as open data. | Delhi (289 wards), Chandigarh (63 sectors, 28 wards) | Live |
| Roads | The main road network, for context. This is reference data, not a hazard. | 25,956 roads in 4 cities | Live |
| Railways | The rail and metro network, for context. This is reference data, not a hazard. | 7,436 rail lines in 4 cities | Live |
| Key facilities | Hospitals, schools, fire stations and police stations. Use these to see what sits inside a hazard area. This is reference data, not a hazard. | 8,577 points in 4 cities | Live |
| Groundwater | How fast the water table falls, from 22 years of CGWB well records. Each district has a measured specific yield. A check against GRACE satellite data runs every week. | 11 districts in Punjab and Haryana | Live |
| Climate outlook | What climate models project up to 2030. The map shows 4 extreme-weather measures under 2 warming paths. It shows the full range across 4 models, not a single average. | 9 areas of North India | Live |
| Forest loss | Points where Sentinel-1 radar detected a possible loss of tree cover. None are confirmed against Hansen or GLAD records yet, and for 114 of them the annual Hansen product shows no loss nearby at all, so each point carries a confidence level and the weakest are hidden by default. | 161 candidate points in 9 areas | In progress |
| Drought | How far rainfall is below normal, using the SPI-3 index over 34 years of CHIRPS data. We have not yet tested it against recorded IMD drought declarations. | Ludhiana. More districts to follow. | Live |
| Cloudburst | Valleys where sudden heavy rain can occur. It combines IMERG rainfall data with monthly visitor numbers. We have not yet tested it against recorded cloudburst dates. | 4 Himalayan pilgrim sites | Live |
| Landslide | 517 recorded landslides from NASA COOLR, with date, trigger and size. Next to them is a model built from slope, relief, monsoon rain and distance to roads. It scores 0.85 in Himachal and 0.81 in Uttarakhand on data it did not train on. It ranks where slopes are prone to fail, not when. A test of a rainfall trigger found no skill, so there is no daily landslide forecast. | Himachal and Uttarakhand | Live |
| Air quality | Live air quality, scored with the CPCB National AQI. This is the index reported in India. The map names the pollutant that sets the level. | 26 India cities and 9 Gulf cities | Live |
| Dam overflow alerts | A warning when a reservoir is near full before heavy rain. Six dams have about 6 years of satellite readings. No alerts are live yet. | Dams in Himachal and Uttarakhand | In progress |
How sure are we?
Every reading has a confidence level
Each reading on the map shows how certain we are. It also shows the source of the value. Click any point to see its confidence level. The levels are Severe, High, Moderate and Low. The forecast adds a Rumored level.
We state where a model applies
The heat model works inside a city that we measured. It explains about 40% of the heat pattern there. It does not work in a city that we did not measure. The map therefore shows heat one city at a time. We test the flood model in the same way, on cities that we did not use to train it.
We test every model on data that we did not use to train it. The home page gives each score, including the low ones. The evaluation page gives the full method.
Flood: what the numbers are not
Limits of the flood layer and the 7-day forecast
- Fails on coastal/tidal mechanisms (Mumbai unseen-city ROC-AUC 0.59) — coastal/tidal feature work is scoped into Sprint 2.
- 30m terrain cannot resolve individual curbs/drains.
- Street-waterlogging model has no probability, only a relative ranking; absence of a documented hotspot is not evidence of safety.
- The urban double-bounce RISE that lets us SEE street flooding is more ambiguous than the open-water drop — new construction, vegetation and cross-orbit incidence-angle changes can also raise backscatter — so it is gated to built-up land and a conservative threshold, and the dual-mechanism labels must be re-pulled + re-evaluated before their accuracy is claimed.
- Sentinel-1's 6-12 day revisit misses flash waterlogging that drains within hours; the SAR post-window must overlap standing water for an event to be labelled at all.
- The served flood probability is per-cell climatology conditioned on live rain - each cell's own past flood rate from Sentinel-1 weeks, raised when the 7-day rain forecast is high. It is not a learned model. is_learned stays FALSE. Wider coverage is not better prediction.
- A cell absent from the forecast is a cell with no label history, not a cell that is safe.
- A SAR reading of dry is weak evidence that a cell stayed dry. Sentinel-1 revisits a place every 6 to 12 days and can miss water that drains within hours, so a zero in a cell's history can be a missed flood.
- Widening the cell set widens exposure to the bbox's rural cells, whose climatology is near zero and whose bands will read Low (or Rumored, the band below Low) regardless of conditions.
- The label history behind each rate covers June to September, 2017 to 2024, only. A rate built from monsoon weeks is a monsoon-season rate, not a rate for other months of the year.
- The review queue is now a top-N; the truncated count must be shown or the human is looking at a biased sample of the disagreements.
This list is read from configs/hazard/flood.yaml, the file the flood pipeline reads.
More detail
The data, the models and the accuracy numbers are all open on GitHub.