Nowcasting Meso-γ-Scale Convective Storms Using Convolutional LSTM Models and High-Resolution Radar Observations

As a deep learning approach to improving precipitation nowcasting, this study proposes convolutional Long Short-Term Memory (LSTM) models which specifically target meso-γ-scale, localized convective storms although there are numerous types of thunderstorms across scales. A Convolutional LSTM Model (CLM) and an Encoder-Decoder Model (EDM) were built by employing LSTM networks that perform better with sequential data. For training, radar reflectivity (Z) datasets of the Multi-Parameter Phased Array Weather Radar (MP-PAWR) for a convective storm event that occurred over the Tokyo metropolitan area were used as input to the models. An Advection Forecast Model (AFM) that utilizes an optical flow method to generate u and v motion vectors and the Lagrangian advection scheme was also developed to compare with nowcasts from the LSTM models. Model performances were assessed using statistics and skill score measures such as Critical Success Index (CSI) and Probability of Detection (POD) with lead time up to 10 min. It was found that for a total rain area with Z > 10 dBZ, the CLM had the best skill, showing higher CSI scores and correlation coefficients than other models for all lead times. For a convective rain area with Z > 35 dBZ, the CLM showed equal to or slightly higher CSI scores than the AFM until the lead time of 7.5 min but underpredicted the strength of convective cells with a lower CSI at 10 min. Both the CLM and EDM showed more negative Z biases at longer lead times, resulting in lower CSI and POD scores than the AFM in this convective rain category. The higher skill of the AFM at longer lead times is most likely because the convective cells were advected without changing their shapes and intensities largely in this short period.