DeepSpectralRetrieval

This repo contains the codes for the deep learning-based retrieval of snowfall microphysics from dual-frequency spectral radar measurements. The retrieval is described in the corresponding publication: Billault-Roux et al., 2023, Atmospheric Measurement Techniques, https://doi.org/10.5194/amt-16-911-2023.

Questions should be directed at anne-claire.billault-roux@epfl.ch

Structure

It consists of three directories:

• spectra_database_creation: scripts to generate the synthetic training set (used for training the decoder part of the model).
• decoder: scripts to train the decoder part of the model (once training set is generated)
• encoder : script to train the encoder part of the model. Note that this requires to have pre-processed files containing dual-frequency Doppler spectral radar measurements remapped to a common grid.

Requirements

Generation of the training set with a radiative transfer model

This requires a preliminary installation of PAMTRA (https://pamtra.readthedocs.io/en/latest/).
Note that to this date, PAMTRA does not run in conda environments: scripts in the spectra_database_creation directory should therefore not be run with an active conda environment.

Deep learning framework

For this part however, we recommend that the user sets up a conda environment to use / run the codes of the deep learning framework (i.e. in the decoder and encoder directories).
python3 is required (python3.7 to 3.9 were tested) , with at least the following packages:

pytorch>=1.9
tensorboard>=2.6 (see e.g. https://pytorch.org/docs/stable/tensorboard.html )
scipy>=1.5.3
pandas>=1.3.5
h5py>=3.3
numpy>=1.18
matplotlib>=3.4
tqdm

Examples of data on which to run the code were not included (yet) because quite heavy, but can be shared upon request.