Figure 1

Process flow where the generator predicts radar frames using self-supervised labels. After passing through the stem of the generator the four resolution representations, consisting of a feature extractor containing the convolution layers for each resolution, are fused. The features classify precipitation types from the classifier (section “Clustering methods for self-supervised learning”) and initialize ConvGRU cells with the latent vector according to the type. The specific structure of the network is described in Fig. 3.