Synthetic hail experiments The data used for assessing the performance of drone-based hail photogrammetry. Three types of hail objects were used: EPS (extended polystyrene) Glass pebbles Ice from a consumer-grade ice-maker The experiments were performed on different types of grass, such as soccer fields, lawns and meadows. The models are trained on a leave-one-out cross-validation (LOOCV) scheme, where the test set is based on one surface, while training and validation sets are based on the rest (randomly split). This is defined as an experiment configuration (a-e), where the name refers to the surface of the test set. Code The code used to create and analyze the data can be found on github: https://github.com/MeteoSwiss/ehw24-hail-photogrammetry

data.zip Contains the processed data to create plots eps_experiment.nc ice_experiment.nc glass_experiment.nc detection_scores.nc detection_fp.nc detection_fn.nc experiments.zip Contains the full data that need to be processed for each expert and hail type annotation eps_jportmann_a annotations instances_Test.json instances_Train.json train.json val.json images Test Train eps_jportmann_b ... ... detections eps_jportmann_a pkl eps_jportmann_b ... models eps_jportmann_a events.out.tfevents last_checkpoint metrics.json model_final.pth eps_jportmann_b ... ... additional_plots.zip Loss curves during model training for each expert and hail type