CONDOR (CGAN) Training/Validation dataset – 69 6000 patches of convective scenes selected from collocated MSG-SEVIRI--GPM-DPR data between 2018 and 2025

CONDOR (Continuous cONvective Depth and Organization Reconstruction) is a generative deep learning model based on a conditional generative adversarial that reconstruct a proxy of convective depth and intensity at geostationary spatio-temporal resolution. It produces a continuous monitoring of deep convection through a radar-derived proxy of convective vertical extent: the Echo-Top Height at 20 dBZ (ETH20). Collocated MSG-SEVIRI multispectral infrared imagery and GPM Dual-frequency Precipitation Radar (DPR) observations over tropical Atlantic and Africa are used to train the retrieval of ETH20 from multispectral infrared sequences.
The provided dataset served as a training and validation dataset for the generative model. It contains a selection (cropping) of 69 600 48x48-pixels patches of temporal and spatial collocation of geostationary (MSG-SEVIRI) and radar (GPM-DPR) for the analysis deep convection over Africa and tropical Atlantic. It includes collocated infrared brightness temperatures from MSG-SEVIRI with ETH20 derived from GPM-DPR reflectivity profiles, along with the viewing angle of MSG-SEVIRI. This database were used to train and validate the model. All details are given in Netz et al.: “Continuous mapping of deep convective cores from geostationary infrared imagery using a multi-scale composite-loss conditional GAN trained on GPM-DPR”, preprint that will be submitted in AIES on March 2026.

References:

Netz et al., 2026 in prep. (Submission planned in AIES journal in March 2026, DOI cited within the paper once released)