HailPred: Hail Weather Recognition Based on Spatiotemporal Series Analysis Algorithm

A spatiotemporal series dataset based on S-band Doppler weather radar data is constructed based on hail observation records. Since the hail weather with both temporal and spatial features, the Causal long short-term memory (LSTM) computation unit with Gradient Highway Unit (GHU) and ResNet18 network is initially used to build a spatiotemporal series model named HailPred based on deep learning. The HailPred hail recognition model was trained several times to achieve the best performance. The HailPred model has a recognition accuracy of 96%, a probability of detection (POD) of 97%, a false alarm rate (FAR) of 5%, and a critical success index (CSI) of 93%, which is 12 percent higher than the accuracy of machine learning algorithms. The recognition performance is significantly better than other algorithms and could play an important role in hail disaster mitigation.