Identification of Atmospheric Dynamics of a Tropical Hail Event and Application of Radar Reflectivity Nowcasting Using a UNet-ConvLSTM2D Deep Learning Model

Abstract

Hail is a small-scale hydrometeorological phenomenon threatening disaster-prone regions of Indonesia, yet understanding of its atmospheric dynamics and deep learning-based radar nowcasting remain limited. This study characterizes the atmospheric conditions of the Depok City hailstorm on 28 October 2024 and evaluates a hybrid UNet-ConvLSTM2D model in predicting radar reflectivity. The study employed Column Maximum (CMAX) radar reflectivity data in PNG and VOL formats from the Soekarno-Hatta Class I Meteorological Station, cloud-top temperature and Cloud Convective Overlay data from Himawari-9 satellite imagery, and ERA5 reanalysis data on storm thermodynamic and kinematic conditions, while radar data from January 2023 to October 2024 were used for model training and validation. Model performance was assessed using Mean Squared Error (MSE), Mean Absolute Error (MAE), Structural Similarity Index (SSIM), Critical Success Index (CSI), and Heidke Skill Score (HSS) across five reflectivity thresholds. Results show maximum radar reflectivity exceeding 55 dBZ at 08.18 UTC and 08.42 UTC, cloud-top temperatures of approximately -61°C to -73°C, and Convective Available Potential Energy (CAPE) of 524-590 J/kg with weak vertical wind shear of 2.4-2.8 m/s from ERA5, confirming a moderately unstable tropical convective environment. Prediction results yielded average values of 184.5897 for MSE, 9.7081 for MAE, and 0.3956 for SSIM, with CSI decreasing from 0.588 to 0.284 and HSS from 0.637 to 0.057 as lead time increased from 8 to 56 minutes. These findings are expected to enhance understanding of the atmospheric dynamics contributing to hail and support wider deep learning-based nowcasting for early warning in tropical Indonesia.