Model Card for LeeWaveNet

This repository contains four neural-network models, trained using fastai, for detecting and determining characteristics of trapped lee waves using maps of 700 hPa vertical velocity as input.

• The base model segmodel.pkl generates a segmentation mask indicating where trapped lee waves are present. This model uses a U-Net architecture with Resnet-34 (pre-trained on ImageNet) as the encoder model.
• Three alternative model heads have been trained on synthetic data: amplitude_0.0625.pkl, wavelength_0.125.pkl and orientation_0.25.pkl. These predict the amplitude, wavelength and orientation of detected waves respectively.

For full details, please see the article by Coney et al. (2023).