Convective Initiation Nowcasting in South China using Physics-augmented Random Forest Models and Geostationary Satellites

Convective initiation (CI) nowcasting in subtropical regions, such as South China, is often hindered by complex atmospheric processes and the imbalanced distribution of CI events, leading to high false alarm ratios (FAR). To address these challenges, this study develops a Storm Warning System with Physics-Augmentation (SWASP), which integrates a random forest (RF) algorithm with cloud physical conditions and auxiliary geospatial information. The model leverages multi-channel data from the Himawari-8 Advanced Himawari Imager (H08/AHI) during the warm seasons (April to September) of 2019.

The SWASP model incorporates critical physical indicators associated with CI triggering mechanisms, including cloud-top cooling rates, cloud-top height relative to the tropopause, and the temporal evolution of cloud-top height. Given that the physical thresholds for convective development vary regionally, we recalculated six threshold criteria based on these physical variables, evaluating multiple percentile-based schemes. The most effective performance was achieved when applying the 85th percentile criterion to define pre-convective cloud conditions, which was then used as input to the RF model. This configuration yielded the highest critical success index (CSI) and the lowest FAR compared to traditional threshold-based methods.

In addition to cloud physical parameters, the model also integrates topographic elevation, satellite zenith angle (SAZ), and latitude (LAT) to account for regional and observational variability. Compared with conventional methods, the SWASP model improves the probability of detection (POD) by 0.11 and 0.08, while reducing FAR by 0.38 and 0.44 during daytime and nighttime, respectively. Moreover, the system demonstrates the capability to detect local convective storm systems approximately 30 minutes to 1 hour ahead of radar-based detection in typical CI cases.

This study highlights the advantage of integrating physical knowledge into machine learning-based nowcasting frameworks and demonstrates the potential of geostationary satellite observations in enabling timely and accurate convective early warnings in subtropical regions.