Detection of bow echoes in French kilometer-scale models (AROME-EPS & AROME models of Météo-France)

To detect bow echoes (BE) directly in simulated reflectivity fields from French kilometer-scale models, three datasets are available. These datasets allowed for fitting and testing a univariate U-Net (convolutional neural network) using simulated reflectivity fields.

Grid area : 717x1121 grid points, Western Europe (12°W-16°E and 37.5°N-55.4°N, resolution 0.025°)

In each tar.gz file, files named 'input_BE' are reflectivity fields from operational models (in mm per hour), 'groundTruth' files correspond to manually labeled contours of BE (field with 1 for each grid point in BE and 0 for outside). The file format is HDF5. HDF5 files contain an array of N fields (array shape : (N,1,717,1121))

- dataset_train.tar.gz (N=6206): input and target fields to train the U-Net (only from the AROME-EPS model)

- dataset_validation.tar.gz (N=2620): input and target fields to validate the U-Net (only from the AROME-EPS model). Training and validation databases contain independent weather case studies.

- dataset_det_AROME.tar.gz  (N=348): input and target fields to apply U-Net to the French deterministic AROME model (same grid than one of AROME-EPS).  Weather case studies are the same than those in the validation database.

- optimal_UNet_architecture.json : architecture of the optimal U-Net configuration

- optimal_UNet_weights.h5 : weights of the trained optimal U-Net configuration