gridded-obs

gridded_obs a package for the verification of gridded observations

• Any quantity, does not care about units

• Easy to add readers

• Contingency scores

• Lmin from Fraction Skill Score

• Correlation coefficient

Like other verification packages, Obtaining score figures with gridded_obs is a two-step process. 1. Compute atomic scores for each forecast that will be verified 2. Aggregate scores for a given period, lead time, start hour, etc.

Getting the code

To get gridded_obs, simply clone this repository.

Add gridded_obs to your working conda environment

• Activate a conda environment with a Python installation.

  If you don't already have one, you can build one with:

  conda create -n gridded_obs_test_environment  
  

  Here, gridded_obs_test_environment is the name of the environment we are creating. You can chose any other name.

• Install gridded_obs and domcmc

  conda install gridded_obs domcmc -c dja001
  

  If you are not using "standard" files used at the CMC in Dorval, you can omit to install domcmc.

  ALTERNATIVELY,

  If you want make modifications to gridded_obs, you can install the editable package.

  Install dependencies manually:

  conda install dask domutils domcmc -c dja001
  

  Install an editable version of gridded_obs:

  pip install --editable /path/to/gridded_obs/package
  

1- Compute atomic scores

Atomic scores can be computed from an interactive session on a compute node.

• Start by requesting a compute node on one of the PPPs

  qsub -I -lselect=1:ncpus=80:mem=185gb,place=scatter:excl -lwalltime=6:0:0
  

• Activate the conda environment that allows to run gridded_obs

  First get access to conda with

  eval "$(/fs/ssm/main/opt/intelcomp/master/inteloneapi_2022.1.2_multi/oneapi/intelpython/python3.9/bin/conda shell.bash hook)"
  
  

  Ignore the error: -bash: syntax error near unexpected token '(', I don't know why it shows up and things still work...

  Then activate your gridded_obs environment

  conda activate gridded_obs_test_environment
  

  If this works, you should see (gridded_obs_test_environment) at the beginning of your shell prompt.

• Make a local copy of the launch script that you will be using

  cp .../gridded_obs/scripts/launch_verification.sh ./your_launch_script.sh
  

  The script name is not important but its helpful to relate it to a given project so that the verification can be reproduced later if needed.

• Edit your_launch_script.sh for the experiments and date range that you want to compare. Most default option should be good as a start.

• Compute scores

  ./your_launch_script.sh
  

  Say you verify precipitation every 10 minutes for 12 hours = 96 lead times in total. You can expect the verification to take approx one minute.

• View images

  The images have been generated in the output directory specified by the figure_dir option in the launch script. I like to use firefox to look at them in my public_html.

2- Aggregate scores and generate images

Aggregating scores and making figures is easy and takes no time.

• Make a local copy of the launch script that you will be using

  cp .../gridded_obs/scripts/launch_aggregate.sh ./your_aggregate_script.sh
  

  and edit it for the experiments, period and scores that you want to plot.

• Run it in your conda environment

  ./your_aggregate_script.sh
  

  The figures will be in a directory listing all experiments being verified.

Use multiple compute nodes to accelerate computation of scores

If you are verifying a large number of forecasts, computing sores can take a while. To accelerate this, we can trow more resources at the problem.

Open an interactive session with 10 compute nodes:

qsub -I -lselect=80:ncpus=10:mpiprocs=10:mem=23gb,place=scatter:excl -lwalltime=6:0:0
# activate your conda env with dask
conda activate gridded_obs_test_environment

Start a dask cluster that will use the 10 nodes. This can take a little while.

. ssmuse-sh -x main/opt/intelcomp/master/inteloneapi-mpi_2022.1.2_all
. ssmuse-sh -d /home/mde000/ssm/maestro-dask-cluster/0.6
start_dask_cluster

Compute scores 10x faster than before

./your_launch_script.sh

Stop cluster when you are done using it

stop_dask_cluster