Voodoonet model used in Cloudnet categorize and classification products from expedition ARTofMELT, Arctic Ocean, 2023

Heather Guy, Ian Brooks

This is a convolutional neural network (CNN) model used by the voodoo algorithm to determine the pseudo-probability of liquid water presence from radar doppler spectra.

This model was trained specifically for improved performance when applied to data from the Radiometer Physics RPG-FMCW-94 radar on the Swedish icebreaker Oden during the expedition ARTofMELT (Atmospheric Rivers and the Onset of Arctic Melt).

AtmosphereObservationsArcticArctic OceanARTofMELTARTofMELT 2023 expeditionIcebreaker OdenCloudnetProfilesCloudsCategorizeClassificationVoodooRadar

Name

Version

Citation

References

Tukiainen S, O’Connor E, Korpinen A (2020) CloudnetPy: A Python package for processing cloud remote sensing data. Journal of Open Source Software 5:2123. https://doi.org/10.21105/joss.02123

Schimmel W, Kalesse-Los H, Maahn M, Vogl T, Foth A, Garfias PS, Seifert P (2022) Identifying cloud droplets beyond lidar attenuation from vertically pointing cloud radar observations using artificial neural networks. Atmospheric Measurement Techniques 15:5343–5366. https://doi.org/10.5194/amt-15-5343-2022

Data description

This dataset contains a single PyTorch (pt) file, that is a multi-dimensional array containing the model weights and architecture of the voodoo model RPG_merged_cloudy_resampledv_trained-model.pt that was used to determine the pseudo-probability of liquid water presence from the cloud radar doppler spectra during the ARTofMELT expedition.

The model is 11.4 MB in size and be used directly with the voodoo (Schimmel et al. 2022) and voodoonet algorithm, where voodoonet is the ACTRIS-Cloudnet implementation of voodoo (Tukianinen et al. 2020).

Comments

Algorithm details

Except where specified below, all model architecture and training parameters used are the defaults as defined in Voodoonet version 0.1.10.

This CCN model differs from the default trained CCN model included in the Voodoonet algorithm in two major ways:

1. The input data used to train the model comes from a wider range of ACTRIS-Cloudnet stations that include stations more representative of high latitude environments, as well as RPG cloud radars using different chirp settings.

2. Prior to training, all doppler spectra are resampled on to the same doppler velocity scale. The purpose of this is to reduce the sensitivity of the voodoo algorithm to differences in radar chirp settings, it allows voodoo to use the absolute values of the doppler velocity as part of the phase classification, as opposed to just the shape of the spectra. Note that this may not be appropriate in cases where the doppler velocity of is biased, such as in the case of gravity waves or if strong horizonal winds are aliased into the vertical signal.

Training data

RPG_merged_cloudy_resampledv_trained-model.pt was trained on data from 5 Cloudnet sites (Leipzig, Burcharest, Neumayer, Norunda, and Hyytiala). 150 random days of data were downloaded from each site using the Cloudnet API and filtered so that only days that were cloudy > 40% of the time where included in the training data, this resulted in a total of 596 days of data (~120 days from each site, roughly evenly distributed between seasons).

Associated raw data and code

Voodoonet version 0.1.10 was obtained from https://github.com/actris-cloudnet/voodoonet.

The data used to train this model are generated by the Aerosol, Clouds and Trace Gases Research Infrastructure (ACTRIS). We acknowledge ACTRIS and Finnish Meteorological Institute for providing the data set which is available for download from https://cloudnet.fmi.fi. We acknowledge ECMWF for providing IFS model data as part of the generation of these ARCTRIS datasets.

References to data used to train the model:

Moisseev D, O'Connor E, Petäjä T (2025) Categorize data from Hyytiälä. ACTRIS Cloud remote sensing data centre unit (CLU). https://cloudnet.fmi.fi/site/hyytiala

Eliasson S, Mölder M, O'Connor E (2025) Categorize data from Norunda. ACTRIS Cloud remote sensing data centre unit (CLU). https://cloudnet.fmi.fi/site/norunda

Adam M, Pîrloagă R, Țoancă F, O'Connor E, Nemuc A, Vasilescu J (2025) Categorize data from Bucharest. ACTRIS Cloud remote sensing data centre unit (CLU). https://cloudnet.fmi.fi/site/bucharest

Engelmann R, Schmithüsen H, O'Connor E, Radenz M (2025) Categorize data from Neumayer Station. ACTRIS Cloud remote sensing data centre unit (CLU). https://cloudnet.fmi.fi/site/neumayer

Foth A, Kalesse-Los H, O'Connor E (2025) Categorize data from Leipzig LIM. ACTRIS Cloud remote sensing data centre unit (CLU). https://cloudnet.fmi.fi/site/leipzig

Project

ARTofMELT 2023 expedition

Publisher

Bolin Centre Database

License

First name

Heather

Last name or organisation

Guy

Email address

heather.guy@ncas.ac.uk

Address

National Centre for Atmospheric Science (NCAS); Fairbairn House; 71-75 Clarendon Rd

Postal code

LS2 9PH

City

Leeds

Province or state

Yorkshire

Country

United Kingdom

GCMD science keyword

Earth science

GCMD location

Ocean > Arctic Ocean

Dataset language

English

DOI

Time