Google/DeepMind AI riverine flood forecasting (Flood Hub)

entry · p2026-09-06-294 · validated

This enables us to provide critical flood information to 700 million people worldwide, a considerable leap from our previous reach of 460 million.

tech: learned-flood-forecasting
passes the definition as: an LSTM model fit to global streamflow and weather records that predicts river discharge in watersheds with no gauge at all, in place of a hydrological model that must be calibrated to a long record in each basin
from: 2023
region: world
domain: safety
side: benefit
people: 700,000,000 people (2024-11)
effect: 5 days of lead time (2024-03) — reliable warning of extreme riverine floods in ungauged basins at up to five days ahead, where the previous state of the art (Copernicus GloFAS) was that reliable only at zero-day lead; the operator reports the same accuracy at seven days as of November 2024
reversible: yes
who holds the switch: owner:Google
attribution: named
source kind: primary
the statistic: Nearing et al., 'Global prediction of extreme floods in ungauged watersheds', Nature 627, 559-563 (March 2024): the model 'achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time that is similar to or better than the reliability of nowcasts (zero-day lead time) from a current state-of-the-art global modelling system', and additionally matches at five-year return periods what was previously achieved at one-year return periods (https://www.nature.com/articles/s41586-024-07145-1). Google reports the operational reach on 11 November 2024. NOTE THE WEAKNESS HONESTLY: 700 million is coverage, not people warned; it is the same shape as the Maps ETA entry already in the ledger (users times an accuracy gain) and the same shape the previous harvest declined to file for ECMWF AIFS. Per person the benefit is ordinary until a flood arrives and then absorbing; there is no source that counts people actually alerted or deaths avoided, which is what would settle the row. The paper cites the general finding that early warning systems cut flood fatalities by up to 43 percent, but attributes no share of that to this model.
source: https://blog.google/innovation-and-ai/products/expanding-flood-forecasting-coverage-helping-partners/
proposed by: researcher (opus), brief 2026-09-06, severe-benefits
received: 2026-09-06T17:40:15-04:00
harvest report: impacts/harvest/2026-09-06-severe-benefits.md

Rulings, in order

2026-09-06T17:40:16-04:00 · validatorpass — quote found on the fetched page (exact match; HTTP 200 text/html; charset=utf-8)
2026-09-06T17:40:46-04:00 · sessionseverity — Flood warnings up to five days ahead available to 700 million people: coverage, not people warned, and ordinary until a flood comes; the absorbing benefit (lives) has no count, as the researcher says. Ordinary at the coverage count (2024); re-rule the day someone counts the warned or the saved.

The proposal as it arrived (JSON)

{
 "id": "p2026-09-06-294",
 "received": "2026-09-06T17:40:15-04:00",
 "proposer": "researcher (opus), brief 2026-09-06, severe-benefits",
 "kind": "entry",
 "sha": "3c70ede6ac9ecfe6",
 "data": {
  "kind": "entry",
  "tech": "learned-flood-forecasting",
  "tech_name": "Google/DeepMind AI riverine flood forecasting (Flood Hub)",
  "definition": "an LSTM model fit to global streamflow and weather records that predicts river discharge in watersheds with no gauge at all, in place of a hydrological model that must be calibrated to a long record in each basin",
  "region": "world",
  "domain": "safety",
  "side": "benefit",
  "from": "2023",
  "people": {
   "n": 700000000,
   "unit": "people",
   "date": "2024-11"
  },
  "effect": {
   "n": 5,
   "unit": "days of lead time",
   "date": "2024-03",
   "what": "reliable warning of extreme riverine floods in ungauged basins at up to five days ahead, where the previous state of the art (Copernicus GloFAS) was that reliable only at zero-day lead; the operator reports the same accuracy at seven days as of November 2024"
  },
  "reversible": true,
  "switch": "owner:Google",
  "statistic": "Nearing et al., 'Global prediction of extreme floods in ungauged watersheds', Nature 627, 559-563 (March 2024): the model 'achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a five-day lead time that is similar to or better than the reliability of nowcasts (zero-day lead time) from a current state-of-the-art global modelling system', and additionally matches at five-year return periods what was previously achieved at one-year return periods (https://www.nature.com/articles/s41586-024-07145-1). Google reports the operational reach on 11 November 2024. NOTE THE WEAKNESS HONESTLY: 700 million is coverage, not people warned; it is the same shape as the Maps ETA entry already in the ledger (users times an accuracy gain) and the same shape the previous harvest declined to file for ECMWF AIFS. Per person the benefit is ordinary until a flood arrives and then absorbing; there is no source that counts people actually alerted or deaths avoided, which is what would settle the row. The paper cites the general finding that early warning systems cut flood fatalities by up to 43 percent, but attributes no share of that to this model.",
  "source": "https://blog.google/innovation-and-ai/products/expanding-flood-forecasting-coverage-helping-partners/",
  "quote": "This enables us to provide critical flood information to 700 million people worldwide, a considerable leap from our previous reach of 460 million.",
  "attribution": "named",
  "source_kind": "primary"
 },
 "from_file": "impacts/harvest/2026-09-06-severe-benefits.md"
}