Advancing the Prediction of Severe Thunderstorms Over the Eastern Parts of India Using Satellite Observations and High-Resolution Models

Abstract

Thunderstorms are extreme weather events that result from severe convection, often associated with heavy rainfall, lightning, hail, gusty winds, squall lines, and sometimes even tornados. The pre-monsoon (March to May) season is significant for studying thunderstorms, as most Indian regions, especially the eastern part, experience frequent thunderstorms during this season. Each year, thunderstorms cause heavy damage to lives, properties, and livelihoods in this region. Hence, understanding the nature of thunderstorms in India is a focus of researchers around the country, which could, in turn, lead to better prediction of thunderstorm events. In this regard, numerical weather prediction (NWP) is an essential component of early warning of the occurrence of thunderstorms. However, the performance of the NWP models is highly dependent on the accurate representation of the initial conditions. The present thesis takes the opportunity to study ways to improve the prediction of thunderstorms in eastern India based on observation and modeling efforts. Thunderstorms' low spatial and temporal extent make it challenging to observe in a vast country like India, where not enough Doppler weather radar or in situ observation stations are available. It hampers understanding of thunderstorm attributes' spatial and long-term variation in thunderstorm-prone regions. Therefore, this thesis's first working chapter aims to prepare a long-term thunderstorm database for the pre-monsoon season in India using high-resolution satellite rainfall data obtained from Global Precipitation Measurement (GPM). The high-resolution Integrated Multi-Satellite Retrievals for GPM (IMERG) final product is used to identify thunderstorms from 2001 to 2021. The 93rd percentile of rainfall appears to be the optimum threshold for the detection, with a success ratio of 82% (642 events are confirmed out of 786 detected). The derived thunderstorm database enables the study of spatial and temporal variations of thunderstorms in the region. The study highlights an increasing trend of thunderstorm activity in the eastern parts of the country, focusing the attention of the rest of the thesis on the eastern part of India. Different thermodynamic stability indices often describe the favourable atmospheric conditions for thunderstorms. These indices can provide insights into the characteristics of individual thunderstorms. Hence, the first working chapter also brings out the combined use of radiosonde, reanalysis, and satellite data over eastern India to better understand thermodynamic indices during the pre-monsoon season. In this chapter, it has been demonstrated that the thermodynamic indices derived from the INSAT-3D satellite can reliably capture the instability of the atmosphere during thunderstorm days. The second working chapter tries to understand the role of horizontal resolution and downscaling approach for thunderstorm simulation in the Weather Research and Forecasting (WRF) model. The model is configured with two nested domains with 9 km and 3 km domains (DD3), 6 km and 2 km domains (DD2), and a single domain at 3 km resolution (SD3). The average mean errors of 2-m temperature (T2) and 2-m relative humidity (RH2) in the DD2 experiment are 0.7 ̊C and -6%, respectively, at the mature stage and 0.2 ̊C and -4%, respectively, at the dissipating stage. The error in SD3 and DD3 is relatively higher (9-17% for T2 and 20-60% for RH2) relative to DD2. The DD2 could show slightly higher instability (convective available potential energy, CAPE, 3188 J kg-1) as compared with DD3 (3164 J kg-1) and SD3 (3020 J kg-1). The rainfall timing and magnitude have also improved in 8 and 12 cases, respectively, in the DD2 run. A high critical success index and less RMSE of different rainfall thresholds suggest better simulation of rainfall magnitude in the DD2 run. The results highlight that high resolution with nested configuration yields better simulation skills than the single domain configuration at high resolution. The third working chapter investigates the sensitivity of land use land cover (LULC) initialization in the WRF model simulation of thunderstorms. Three types of land use land cover (LULC) maps have been prepared using supervised machine learning methods such as Classification and Regression Trees (CART), Naive Bayes (NB), and Support Vector Machine (SVM) from Landsat 8 data. A high accuracy score (85%) and kappa coefficient (81%) revealed the best performance of the CART classifier in generating the LULC maps. Model results highlight that the CART experiment exhibits relatively less bias in RH2 (~ -10% to -5%), T2 (<2.5°C), and 2m wind speed (-1 to ~1.8 m s-1). The CART could improve the rainfall with the least error (~ -16 mm) compared to CNTL (~ -33 mm), NB (~ -37 mm), and SVM (~ -38 mm), and supported by the quantitative statistical analysis viz., less false alarm ratio, high detection rate and critical success index for all thresholds. LULC class-wise analysis indicates a higher variation of surface and lower atmospheric parameters over urban, shrubland, and cropland while less variation over barren, forest, and water. The fourth working chapter assesses the thunderstorm predictability by assimilation of INSAT 3D atmospheric profiles into the WRF model initial condition through the 3-dimensional variational data assimilation technique. Two different numerical experiments were conducted, EXP-CN (without assimilation) and EXP-IN (with INSAT-derived profiles assimilation). The results show consistently enhanced performance in the representation and simulation of surface variables such as RH2, T2, and WS10 in EXP-IN compared to EXP-CN. The profiles of RH, Winds, and vertical velocity have improved in the EXP-IN experiment. The higher CSI and lowest false alarm ratio (FAR) make it evident that the assimilation of INSAT data enhances the skill in rainfall prediction in thunderstorm study. The outcome of the thesis is that the prediction of thunderstorms and their stages can be significantly improved by 10-17% by configuring the model at high resolution with nested domain, by 10-18% by representing realistic LULC forcing, and by 30-37% by assimilating atmospheric profiles of temperature and moisture.