Prediction of thunderstorm evolution using deep learning models with doppler weather radar observations over southern part of India

Thunderstorms are severe weather phenomena causing heavy rainfall and lightning that poses serious threat to agriculture, infrastructure and lives in general. Short scale nature of these events makes it difficult to predict them. In this study we have made an attempt at thunderstorm nowcasting using Deep Learning (DL) models with Doppler weather radar (DWR) data, for the first time over the Indian region. This study utilizes data from a C-band DWR installed at Space Physics Laboratory (8.52 N, 76.89E), Thiruvananthapuram (southern tip of India). DL models incorporating Generative Adversarial Network (GAN) architecture have been developed to predict evolution of pre-monsoon (Mar-May) thunderstorms over southern peninsular India. MAXZ (maximum reflectivity along the vertical) reflectivity data of thunderstorm events during pre-monsoons of 2018–2024 have been used for training and testing of the DL models. Total 4 models have been trained. For 15 & 30 minutes ahead predictions, the Mean Absolute Error (MAE) for the test samples are about 0.8 dB and 1.2 dB respectively. Our DL models are capable of predicting the main convective areas (Z > 40 dBZ) for both 15 and 30 minutes ahead predictions better than some earlier studies. Evolution of the thunderstorm on 13-May-2018 has been studied in detail. Movement of the system has been tracked by following the center of the largest cluster of high reflectivity (Z > 40 dBZ) values. All the four models were able to capture the overall spatial patterns of the reflectivity field well. For 15 minutes ahead predictions, the models predict the movement of the center reasonably well. The scatter plot between the direction of true movement & predicted movement of the center are well correlated. The study demonstrates the ability of the deep learning models in predicting the evolution of thunderstorms over Indian region.