L-CNN model development and analysis

This repository contains the source code for the article Advection-free Convolutional Neural Network for Convective Rainfall Nowcasting by Jenna Ritvanen, Bent Harnist, Miguel Aldana, Terhi Mäkinen, and Seppo Pulkkinen, published in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.

For documentation of the L-CNN method, refer to the article. For questions related to the code, contact Jenna Ritvanen (jenna.ritvanen[at]fmi.fi).

The workflow for replicating the results in the article is roughly:

1. Create the training, validation and test datasets in Lagrangian coordinates.
2. Train the L-CNN model.
3. Create nowcasts for the L-CNN model.
4. Create the reference model nowcasts
   1. Train the RainNet model and run nowcasts for it. See the accompanying repository fmidev/rainnet for details.
   2. Create nowcasts for the LINDA and extrapolation nowcast models. See verification/replication.md for instructions and config/p25-extrap-linda-whole-domain-optical-flow for configuration files.
5. Run verification statistics computation and visualize the results.

Scripts

| Script | Description | | ------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | transform_fmi_composite_to_lagrangian.py | Script for transforming FMI radar composite to Lagrangian coordinates. Reads parameters from lagrangian_transform_datasets.yaml and lagrangian_transform_params.yaml configuration files. | | train_model.py | Script for training the L-CNN model. Dataset parameters read from lagrangian_datasets.yaml and model parameters from lcnn.yaml. | | predict_model.py | Script for producing predictions from the datelists/fmi_rainy_days_bbox_predict.txt datelist. Dataset parameters read from lagrangian_datasets.yaml and model parameters from lcnn.yaml. | | plot_example_nowcasts.py | Script for plotting example nowcasts. Input parameters nowcast time and config file name. | | plot_example_nowcasts_gif.py | Script for plotting example nowcasts as GIF animations. Input parameters nowcast time and config file name. Config file has same structure as plot_example_nowcasts.py, but output filename and figure size are not respected. | | create_nowcast.py | Script for creating a nowcast for a single timestep with the L-CNN model. Input parameters are the timestep, configuration sub-directory, path to the model checkpoint, number of workers for data transform, and path to the datelist file that is written. |

Creating Lagrangian datasets

The Lagrangian-transformed datasets are created with the script transform_fmi_composite_to_lagrangian.py. The configuration for the transformation is given in two configuration files lagrangian_transform_params.yaml for the transformation parameters and lagrangian_transform_datasets.yaml for the dataset configuration.

The script is run with

bash python transform_fmi_composite_to_lagrangian.py <config-sub-path> <dataset-split> --nworkers <workers>

where <config-sub-path> is a sub-directory of the config directory where the configuration files are located, <dataset-split> is the name of the dataset to be generated (that will be injected to the {split} placeholder in the date_list variable in lagrangian_transform_datasets.yaml), and <workers> indicates the number of dask processes used to run the transformation.

Training the L-CNN model

The L-CNN model is trained with the train_model.py script. The script is going to read 5 configuration files:

• lcnn.yaml for the model and training parameters
• lagrangian_datasets.yaml for the dataset parameters
• log_nowcast_callback.yaml for logging parameters
• nowcast_metrics_callback.yaml for verification metrics calculated during training
• output.yaml for random logging outputs

The training will automatically use the dataset list with train injected to {split} placeholder in the date_list variable in lagrangian_datasets.yaml.

In SLURM-based machines, the training can be run with e.g.

```bash

!/bin/bash

SBATCH --account=

SBATCH --job-name=lcnndiffrmse_30lt

SBATCH --partition=gpu

SBATCH --nodes=1

SBATCH --ntasks=1

SBATCH --cpus-per-task=10

SBATCH --mem=32G

SBATCH --time=72:00:00

SBATCH --gres=gpu:v100:1

SBATCH --output=output_%j.txt

SBATCH --error=errors_%j.txt

SBATCH --mail-type=BEGIN,END,FAIL

module load pytorch

srun python train_model.py &> lcnn-train-rmse-30lt.out

seff $SLURM_JOBID ```

where <config-sub-path> is a sub-directory of the config directory where the configuration files are located.

Note that the packages in environment.yml should be installed, if they are not installed in the module environment.

The training will logged in Tensorboard logger that can be viewed through the dashboard. The logs are located at logs/train_{configpath}.

Creating nowcasts for the L-CNN model

The L-CNN model is run with the predict_model.py script. The script is going to read 5 configuration files:

• lcnn.yaml for the model and training parameters
• lagrangian_datasets.yaml for the dataset parameters
• log_nowcast_callback.yaml for logging parameters
• nowcast_metrics_callback.yaml for verification metrics calculated during training
• output.yaml for random logging outputs

In SLURM-based machines, the prediction can be run with e.g.

```bash

!/bin/bash

SBATCH --account=

SBATCH --job-name=lcnndiffrmse30ltpred

SBATCH --partition=fmi

SBATCH --nodes=1

SBATCH --ntasks=1

SBATCH --cpus-per-task=20

SBATCH --mem=32G

SBATCH --time=96:00:00

SBATCH --output=output_%j.txt

SBATCH --error=errors_%j.txt

SBATCH --mail-type=BEGIN,END,FAIL

module load pytorch

srun python predict_model.py .ckpt -l &> lcnn-rmse-pred.out

seff $SLURM_JOBID ```

where <config-sub-path> is a sub-directory of the config directory where the configuration files are located, and <dataset-split> is the name of the dataset to be run (that will be injected to the {split} placeholder in the date_list variable in lagrangian_datasets.yaml).

Running verification results

The verification results need to be run in the verification subfolder.

First, calculating the metrics is run with the verification/scripts/calculate_metrics.py script using the configuration file calculate_metrics.yaml.

Second, the results are visualized with the verification/scripts/plot.py script using the configuration file plot.yaml.

Note that the name of these configuration files does not matter, as the full path is given to the script.

An example SLURM job for calculating the metrics is given below. Note that in this example the calculate_metrics.py script has been copied to the verification directory to solve some issue with paths in the SLURM environment. When running the script on regular systems, this was not necessary.

```bash

!/bin/bash

SBATCH --account=

SBATCH --job-name=lcnnrmse30lt_verif

SBATCH --partition=fmi

SBATCH --nodes=1

SBATCH --ntasks=1

SBATCH --cpus-per-task=20

SBATCH --mem=64G

SBATCH --time=96:00:00

SBATCH --output=output_%j.txt

SBATCH --error=errors_%j.txt

SBATCH --mail-type=BEGIN,END,FAIL

cd verification

module load pytorch

export PYTHONPATH="${PYTHONPATH}:$(pwd -P)"

srun python calculatemetrics.py /calculatemetrics.yaml &> lcnn-mssim-30lt-verif.out

seff $SLURM_JOBID ```

where <config-path>/calculate_metrics.yaml is the full path to the configuration file.

Finally, the results can be visualized with

bash python scripts/plot.py <config-path>/plot.yaml

where <config-path>/plot.yaml is the full path to the configuration file.

Plotting example nowcasts

Case figures of nowcasts can be plotted with

bash python plot_example_nowcasts.py plot_example_nowcasts.yaml <YYYYMMDDHHMM>

and GIF animations with

bash python plot_example_nowcasts_gif.py plot_erad_gifs.yaml <YYYYMMDDHHMM>

where the date placeholder should be replaced with the desired time. Note that the configuration files are the same, but the GIF script does not respect the output filename and figure size parameters.