Machine Learning Prediction Importances

  1. preciprate (0.0586448032508)
  2. rh (0.0429404800655)
  3. cpreciprate (0.0413386989491)
  4. 400omega (0.0402492140743)
  5. 850q (0.0398999828429)

Forecast valid 2022-June-25 12 UTC

USGS Gauges

These points represent river gauges in the United States where flooding is currently being observed.

Exceeds action stage Exceeds minor stage Exceeds moderate stage Exceeds major stage

RiverWatch Gauges

These represent satellite observation locations in the global RiverWatch system where flooding is being observed. RiverWatch is produced by the Dartmouth Flood Observatory

Exceeds 1.33 year RP Exceeds 5 year RP

Machine Learning Predictions

These points are locations where a random forest prediction model (predictors = selected model fields from the GFS analysis; predictand = U.S. National Weather Service reports of flash flooding from 2006-2013) is applied to the GFS weather model output. In this implementation, the model is trained over the U.S. but applied globally, which is why there tend to be a lot of flooding forecasts in the tropics. The model is most applicable in the mid-latitudes, where high-moisture high-instability environments are relatively rarer than in the tropics. The skill of the model predictions will improve as observed floods from tropical regions are added into the prediction matrix. The 5 most important predictors (Gini criteria) in the model are displayed above.

Predicted 5-10% Predicted 10-25% Predicted >25%

Unit Streamflow

This product consists of the streamflow (m^3/s) at each grid cell scaled by the drainage area contributing to that grid cell (km^2). The streamflow is the output of a run of the EF5 hydrologic modeling framework; the hydrologic model is forced by rainfall estimates from NASA’s GPM IMERG (Global Precipitation Measurement Integrated Multi-Satellite Retrievals for GPM) product. IMERG is available at 0.1 deg resolution every 30 minutes north of 60 deg S and south of 60 deg N latitude with a six-hour latency. EF5 is run at 5-km global resolution every 30 minutes as new IMERG precipitation grids come available. This product is useful to diagnose areas where surface flows are high relative to the ability of river and stream networks to contain those flows.


This product is similar to the Unit Streamflow product, but without the result being scaled by the contributing drainage area of each pixel. The units of this product are therefore (m^3/s). Streamflow is useful in areas where the user is familiar with typical flows for that area.

Streamflow Threshold Exceedance

Various geomorphological and hydro-climatological parameters can be used to predict the flows required to exceed bankfull levels at each gridcell. This product is based on a logistic regression model originally developed in the United States and applied to gauged locations there. The logistic regression model has been extrapolated planet-wide for this product. Purple colors represent locations where the equivalent of the U.S. National Weather Service’s “major flood” stage criteria are being exceeded. Red represents “moderate flood”, orange represents “minor flood”, and yellow represents “action stage”.

Hydrologic Model Soil Moisture

This product is the soil moisture state from the same model runs used to generate the Unit Streamflow and Streamflow products. Soil moisture is displayed as the percent of saturation of the near-surface soil layer in the hydrologic model.

Streamflow Return Period

This product compares the current predicted streamflow at each grid cell to a historical archive of simulated streamflows at the same locations. The historical streamflows are reorganized into a Log-Pearson Type III distribution. Then the current predicted streamflow is compared to that distribution and the reciprocal of the annual probability (also known as the “return period” in years) of that particular streamflow value is displayed. This product is not available south of 50 deg S or north of 50 deg N because it relies upon historical simulations, which are forced by historical rainfall estimates from the NASA TRMM mission, and those are only available in tropical and middle latitudes.

Hydrologic Model GPM IMERG Rainfall

This is the GPM IMERG precipitation product displayed as it is used in the hydrologic model.

Hydrologic Model GFS Forecast Rainfall

This is the 72-hour quantitative precipitation forecast (QPF) in mm from the Global Forecast System (GFS) weather model. This precipitation is fed into the hydrologic model, along with the observed rainfall from IMERG, to produce the hydrologic model forecasts of streamflow and soil moisture.

GFS 500hPa Temp

This is the analyzed 500-hPa temperature in deg C from the GFS weather model.

GFS 500hPa Omega

This is the analyzed vertical velocity in Pa/s at the 500-hPa level in the GFS weather model.

GFS 700hPa Relative Humidity

This is the analyzed relative humidity in percent at the 700-hPa level from the GFS weather model.

GFS Convective Precip Rate

This is the 3-hr GFS forecast of the convective precipitation rate in mm/s.

GFS Total Precip Rate

This is the 3-hr GFS forecast of the total precipitation rate in mm/s.

GFS Precipitable Water

This is the analyzed surface to 300-hPa precipitable water in mm from the GFS weather model.

GFS Relative Humidity

This is the analyzed column-integrated relative humidity in percent from the GFS weather model.

GFS Surface Temperature

This is the analyzed surface temperature in deg C from the GFS weather model.

GFS Surface CAPE

This is the analyzed surface-based convective available potential energy in J/kg from the GFS weather model.

GFS Best 4-Layer Lifted Index

This is the analyzed most unstable lifted index (temperature of air parcel lifted adiabatically to a level minus the environmental temperature at that level) taken from 4 different pressure levels in the GFS weather model.

GFS Soil Moisture

This is the analyzed surface to 10-cm below ground fractional volumetric soil moisture from the GFS weather model.

Open Data

GFS analyses and forecasts:
USGS gauge data:
NWS flash flood events used to train random forest model:
Global RiverWatch flood reports:
Historical TRMM data for generating streamflow return periods:
Global topographical data for hydrologic model:
Global datasets for generating a priori hydrologic model parameters.

Open Software

Hydrologic model (EF5) software:
Libraries for generating random forest model:
Underlying map layers and API: and
Image production and processing: