Verif is a command-line tool that lets you verify the quality of weather forecasts for point locations. It can also compare forecasts from different forecasting systems (that have different models, post-processing methods, etc).
The program reads files with observations and forecasts in a specific format (see "Input files" below). The input files contain information about dates, forecast lead times, and locations such that statistics can be aggregated across different dimensions. To ensure a fair comparison among files, Verif will discard data points where one or more forecast systems have missing forecasts. Since Verif is a command-line tool, it can be used in scripts to automatically create verification figures.
Verif version 1.0 has been released (see "Installation Instruction" below). We welcome suggestions for improvements. Verif is developed by Thomas Nipen, David Siuta, and Tim Chui.
- Deterministic metrics such as MAE, bias, correlation, RMSE (e.g.
-m mae) - Threshold-based metrics such as the false alarm rate, ETS, EDI, Yule's Q (e.g.
-m ets) - Probabilistic metrics such as brier score, PIT-histogram, reliability diagrams (e.g.
-m bs) - Special plots like Taylor diagrams (
-m taylor), quantile-quantile plots (-m qq). - Plot scores as a function of date, lead time, station altitude/lat/longitude (e.g.
-x date) - Show scores on maps (
-type map) - Subset the data by specifying a date range and lat/lon range (
-latrange 58,60) - Export to text (
-type text) - Options to adjust font sizes, label positions, tick marks, legends, etc (
-labfs 14) - Anomaly statistics relative to a baseline like climatology (
-c climfile.txt) - Output to png, jpeg, eps, etc and specify dimensions and resolution (
-f image.png -dpi 300)
For a full list, run Verif without arguments.
Prerequisites
Verif requires NetCDF as well as the python packages numpy, scipy, and matplotlib. The python package mpltoolkits.basemap is optional, but provides a background map when verification scores are plotted on a map. Install the packages as follows:
sudo apt-get install netcdf-bin libnetcdf-dev libhdf5-serial-dev
sudo apt-get install python-setuptools python-numpy python-scipy python-matplotlib python-mpltoolkits.basemapInstalling using pip The easiest is to install the lastest version of Verif using pip:
sudo pip install verifVerif should then be accessible type typing verif on the command-line.
Installing from source Alternatively, to install from source, download the source code of the latest version: https://github.com/WFRT/verif/releases/. Unzip the file and navigate into the extracted folder.
Then install Verif by executing the following inside the extracted folder:
sudo python setup.py installThis will create the executable /usr/local/bin/verif. Add this to your PATH environment
variable if necessary (i.e add export PATH=/usr/local/bin/:$PATH to ~/.bashrc). If you do
not have sudo privileges do:
sudo python setup.py install --userThis will create the executable ~/.local/bin/verif. Add the folder to your PATH environment
variable.
Install NetCDF, numpy, scipy, and matplotlib, and basemap (optionally). Then install Verif by executing the following inside the extracted folder:
sudo python setup.py installVerif will then be installed into /usr/local/share/python/ or where ever your python modules are
installed (Look for "Installing verif script to <some directory>" when installing). Add the folder
to your PATH environment variable, if it is not already added.
Fake data for testing the program is found in ./examples/. There is one "raw" forecast file and
one bias-corrected forecast file (where a Kalman filter has been applied). Here are some example
commands to test out:
verif examples/raw.txt examples/kf.txt -m mae
verif examples/raw.txt examples/kf.txt -m ets
verif examples/raw.txt examples/kf.txt -m taylor
verif examples/raw.txt examples/kf.txt -m error
verif examples/raw.txt examples/kf.txt -m reliability -r 0
verif examples/raw.txt examples/kf.txt -m pithistHere is a list of currently supported metrics. Note that the plots that are possible to make depend on what variables are available in the input files.
| Deterministic | Description |
-m alphaindex |
Alpha index |
-m bias |
Mean error |
-m cmae |
Cube-root mean absolute cubic error |
-m corr |
Pearson correlation between obs and forecast |
-m derror |
Error in distribution of deterministic values |
-m dmb |
Degree of mass balance (mean obs / mean fcst) |
-m ef |
Exceedance fraction: fraction that fcst > obs |
-m fcst |
Average forecast value |
-m kendallcorr |
Kendall correlation |
-m leps |
Linear error in probability space |
-m mae |
Mean of forecasts |
-m mbias |
Multiplicative bias |
-m nsec |
Nash-Sutcliffe efficiency coefficient |
-m obs |
Mean of observations |
-m rankcorr |
Spearman rank correlation |
-m rmse |
Root mean squared error |
-m rmsf |
Root mean squared factor |
-m stderror |
Standard error |
-m within |
Percentage of forecasts that are within some error bound |
| Threshold | Description |
-m a |
Fraction of events that are hits |
-m b |
Fraction of events that are false alarms |
-m baserate |
Climatological frequency |
-m biasfreq |
Numer of forecasts / number of observations |
-m c |
Fraction of events that are misses |
-m d |
Fraction of events that are correct rejections |
-m diff |
Difference between false alarms and misses |
-m dscore |
Generalized discrimination score |
-m edi |
Extremal dependency index |
-m eds |
Extreme dependency score |
-m ets |
Equitable threat score |
-m fa |
False alarm rate |
-m far |
False alarm ratio |
-m fcstrate |
Fractions of forecasts (a + b) |
-m hit |
Hit rate |
-m hss |
Heidke skill score |
-m kss |
Hanssen-Kuiper skill score |
-m lor |
Log odds ratio |
-m miss |
Miss rate |
-m n |
Total cases (a + b + c + d) |
-m or |
Odds ratio |
-m pc |
Proportions correct |
-m sedi |
Symmetric extremal dependency index |
-m seds |
Symmetric extreme dependency score |
-m threat |
Threat score |
-m yulesq |
Yule's Q (odds ratio skill score) |
| Probabilistic | Description |
-m bs |
Brier score |
-m bsrel |
Reliability component of Brier score |
-m bsres |
Resolution component of Brier score |
-m bss |
Brier skill score |
-m bsunc |
Uncertainty component of Brier score |
-m ign0 |
Ignorance of the binary probability based on threshold |
-m marginalratio |
Ratio of marginal probability of obs to that of fcst |
-m pitdev |
Deviation of the PIT histogram |
-m quantilescore |
Quantile score |
-m spherical |
Pherical probabilistic scoring rule |
| Special plots | Description |
-m against |
Plots the determinstic forecasts from each file against each other |
-m change |
Forecast skill (MAE) as a function of change in obs from previous forecast run |
-m cond |
Plots forecasts as a function of obs |
-m discrimination |
Discrimination diagram for a specified threshold |
-m droc |
Receiver operating characteristic for deterministic forecast |
-m droc0 |
Like droc, except don't use different forecast thresholds |
-m drocnorm |
Like droc, except trainsform axes using standard normal distribution |
-m economicvalue |
Economic value for a specified threshold |
-m error |
Decomposition of RMSE into systematic and unsystematic components |
-m freq |
Show frequency distribution of obs and fcst |
-m igncontrib |
Shows how much each probability issued contributes to total ignorance |
-m impact |
Compares two forecast inputs and shows where the improvements come from |
-m invreliability |
Reliability diagram for a specified quantile |
-m marginal |
Marginal distribution for a specified threshold |
-m meteo |
Show forecasts and obs in a meteogram |
-m obsfcst |
A plot showing both obs and fcst |
-m performance |
Diagram showing POD, FAR, bias, and threat score |
-m pithist |
Histogram of PIT values |
-m qq |
Quantile-quantile plot |
-m reliability |
Reliability diagram for a specified threshold |
-m roc |
Receiver operating characteristics plot for a specified threshold |
-m scatter |
A scatter plt of obs and fcst |
-m spreadskill |
Plots forecast spread vs forecast skilL |
-m taylor |
Taylor diagram showing correlation and fcst stdev |
-m timeseries |
Time series of obs and forecasts |
To verify your own forecasts, the easiest option is to put the data into the following format:
# variable: Temperature
# units: $^oC$
date leadtime location lat lon altitude obs fcst p10 q0.1
20150101 0 214 49.2 -122.1 92 3.4 2.1 0.914 -1.9
20150101 1 214 49.2 -122.1 92 4.7 4.2 0.858 0.1
20150101 0 180 50.3 -120.3 150 0.2 -1.2 0.992 -2.1Any lines starting with '#' can be metadata (currently variable: and units: are recognized). After that is a header line that must describe the data columns below. The following attributes are recognized:
- date (in YYYYMMDD)
- unixtime (in seconds since 1970-01-01 00:00:00 +00:00)
- leadtime (forecast lead time in hours)
- location (station identifier)
- lat (in degrees)
- lon (in degrees)
- obs (observations)
- fcst (deterministic forecast)
- p<number> (cumulative probability for a specific threshold, e.g. p10 is the CDF at 10 degrees)
- q<number> (temperature for a specific quantile e.g. q0.1 is the 0.1 quantile)
Either 'date' or 'unixtime' can be supplied. obs and fcst are the only required columns. Note that the file will likely have many rows with repeated values of leadtime/location/lat/lon/altitude. If station and lead time information is missing, then Verif assumes they are all for the same station and lead time. The columns can be in any order.
Deterministic forecasts will only have "obs" and "fcst", however probabilistic forecasts can provide any number of cumulative probabilities. For probabilistic forecasts, "fcst" could represent the ensemble mean (or any other method to reduce the ensemble to a deterministic forecast).
For compatibility reason, 'offset' can be used instead of 'leadtime', 'id instead of 'location', and 'elev' instead of 'altitude'.
For larger datasets, the files in NetCDF are much quicker to read. The following dimensions, variables, and attributes are understood by Verif:
netcdf format {
dimensions:
time = UNLIMITED;
leadtime = 48;
location = 10;
ensemble = 21;
threshold = 11;
quantile = 11;
variables:
int time(time); // Valid time of forecast initialization in
// number of seconds since 1970-01-01 00:00:00 +00:00
float leadtime(leadtime); // Number of hours since forecast init
int location(location); // Id for each station location
float threshold(threshold);
float quantile(quantile); // Numbers between 0 and 1
float lat(location); // Decimal degrees latitude
float lon(location); // Decimal degrees longitude
float altitude(location); // Altitude in meters
float obs(time, leadtime, location); // Observations
float fcst(time, leadtime, location); // Deterministic forecast
float cdf(time, leadtime, location, threshold); // Accumulated prob at threshold
float pdf(time, leadtime, location, threshold); // Probability density at threshold
float x(time, leadtime, location, quantile); // Threshold corresponding to quantile
float pit(time, leadtime, location); // CDF for threshold=observation
// global attributes:
: long_name = "Temperature"; // Used to label axes in plots
: standard_name = "air_temperature"; // NetCDF/CF standard name of the forecast variable
: verif_version = "1.0.0"; // Not required, but will be parsed in the future if format changes
}Copyright © 2015-2017 UBC Weather Forecast Research Team. Verif is licensed under the 3-clause BSD license. See LICENSE file.