Probabilistic Convective Initiation Nowcasting Using Himawari-8 AHI with Explainable Deep Learning Models

Abstract Convective initiation (CI) nowcasting is crucial for reducing losses of human life and property caused by severe convective weather. A novel deep learning method based on the U-net model (named as CIUnet) was developed for forecasting CI during the warm season with eight interest fields of Himawari-8 Advanced Himawari Imager (AHI) and terrain height. The results showed that the CIUnet model produced probability forecasts of CI occurrence location and time with POD (probability of detection) at 93.3±0.3% and FAR (false alarm ratio) at 18.3±0.4% at a lead time of 30-min. Sensitivity and permutation importance experiments on the input fields of the CIUnet model revealed that the differences in brightness temperature for spectral channels were more critical for CI nowcasts than the original infrared channel brightness temperatures. The brightness temperature difference between Band10 (7.3 μm ) and Band13 (10.4 μm ), which represents the cloud-top height relative to the lower-troposphere, is identified as the most important input fields for CI nowcasting. The tri-spectral brightness temperature difference (TTD), which represents cloud-top glaciation, is ranked the second and it significantly reduced the FAR of the CI forecast. Using terrain heights as an extra input feature improved the POD, but slightly overestimated CI over complex terrain. In addition, a layer-wise relevance propagation (LRP) analyses was performed, and confirmed that the CIUnet model can effectively identify the crucial regions and features of the input fields for accurate CI prediction. Therefore, both permutation importance experiments and LPR analyses are useful for improving the CIUnet model and advancing the understanding of CI mechanisms.