VoodooNet

Predicting liquid droplets in mixed-phase clouds beyond lidar attenuation using artificial neural nets and Doppler cloud radar spectra

VOODOO is a machine learning approach based convolutional neural networks (CNN) to relate Doppler spectra morphologies to the presence of (supercooled) liquid cloud droplets in mixed-phase clouds.

Installation

Prerequisites

VoodooNet requires Python 3.10 or newer.

Before installing VoodooNet, install PyTorch according to your infrastructure. Otherwise pip installs the default PyTorch build, which on Linux includes CUDA libraries and is several gigabytes. For example on a Linux machine without GPU you might run:

pip3 install torch --extra-index-url https://download.pytorch.org/whl/cpu

From PyPI

pip3 install voodoonet

To log training runs with Weights & Biases, install the train extra:

pip3 install voodoonet[train]

Locally for development

pip3 install -e .[dev]

Citing

If you wish to acknowledge VoodooNet in your publication, please cite:

Schimmel et al. (2022). Identifying cloud droplets beyond lidar attenuation from vertically pointing cloud radar observations using artificial neural networks. Atmos. Meas. Tech., 15(18), 5343–5366. https://doi.org/10.5194/amt-15-5343-2022

Usage

Make predictions using the default model and settings

import glob
import voodoonet

rpg_files = glob.glob('/path/to/rpg/files/*.LV0')
probability_liquid = voodoonet.infer(rpg_files)

You can for example plot the resulting liquid probability:

import matplotlib.pyplot as plt

plt.pcolor(probability_liquid.T)
plt.show()

Generate a training data set

Download some RPG-FMCW-94 raw files and corresponding classification files from the Cloudnet data portal using cloudnet-api-client, which is installed with voodoonet. For example, for Leipzig LIM on 2021-01-10:

import voodoonet
from cloudnet_api_client import APIClient

client = APIClient()
rpg_meta = client.raw_files(
    site_id="leipzig-lim",
    instrument_id="rpg-fmcw-94",
    filename_suffix=".LV0",
    date="2021-01-10",
)
classification_meta = client.files(
    site_id="leipzig-lim",
    product_id="classification",
    date="2021-01-10",
)
rpg_files = client.download(rpg_meta, "data/")
classification_files = client.download(classification_meta, "data/")
voodoonet.generate_training_data(rpg_files, classification_files, 'training-data-set.pt')

Alternatively, just use N random days from a site. Files are downloaded into download_dir (default cloudnet-data) and reused on subsequent runs:

import voodoonet
voodoonet.generate_training_data_for_cloudnet('leipzig-lim', 'training-data-set.pt', n_days=5)

Train a VoodooNet model

import voodoonet

pre_computed_training_data_set = 'training-data-set.pt'
voodoonet.train(pre_computed_training_data_set, 'trained-model.pt')

Make predictions using the new model

import glob
import voodoonet
from voodoonet.utils import VoodooOptions

rpg_files = glob.glob('/path/to/rpg/files/*.LV0')
options = VoodooOptions(trained_model='trained-model.pt')
probability_liquid = voodoonet.infer(rpg_files, options=options)