Abstract Accurate prediction of atmospheric instability, particularly the lifted index (LI), is critical for anticipating severe convective phenomena such as thunderstorms and lightning over tropical regions. This study develops and rigorously evaluates a satellite-driven, physically interpretable machine-learning (ML) framework for continuous regional-scale LI prediction over Telangana, India, focusing on the pre-monsoon season (April–May) using eleven years (2015–2025) of INSAT-3D/3DR Level-2 sounder observations. District-scale LI variability across ten representative districts is modelled using physically meaningful thermodynamic and dynamical predictors, including Layer-1 Precipitable Water (L1PW), wind index (WI), surface air temperature, surface pressure, and clear-sky brightness temperature at 11 μm. Six state-of-the-art regression algorithms, namely decision tree, random forest (RF), support vector regression, artificial neural network, eXtreme gradient boosting, and light gradient boosting machine (LGBM), are assessed using a temporally independent, year-wise data-partitioning strategy (training: 2015–2021; validation: 2022–2023; testing: 2024–2025) to minimize temporal autocorrelation and ensure operationally realistic generalisation. Model skill is quantified exclusively on the independent test period using multiple complementary performance metrics and a composite skill score. Results show that tree-based ensemble models consistently outperform single learners, with LGBM emerging as the most robust approach, achieving a regional-mean root mean squared error (RMSE) of 2.0 ± 0.15 K and mean coefficient of determination (R2) of 0.72, corresponding to an RMSE improvement of approximately 4% (≈ 0.08 K) and a ΔR2 ≈ +0.04 relative to the RF baseline. SHapley Additive exPlanations (SHAP)-based interpretability analysis identifies L1PW and WI as the dominant physical controls on LI variability, consistent with established theories of tropical convective instability. Independent lightning observations from the World Wide Lightning Location Network (April 2024) show enhanced lightning occurrence during ML-predicted unstable LI regimes, providing an external physical consistency check rather than deterministic lightning prediction. Overall, this study demonstrates the operational feasibility of combining indigenous INSAT sounder observations with physically interpretable ML models for continuous regional instability monitoring, with clear potential for integration into IMD and TSDPS early-warning systems, offering a scalable framework for enhanced thunderstorm early warning over data-sparse regions of India. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Developed a satellite-driven, physically interpretable machine-learning framework for continuous Lifted Index prediction over Telangana using 11 years (2015–2025) of INSAT-3D/3DR sounder observations. </ItemContent> <ItemContent> Implemented a temporally independent, year-wise validation strategy, ensuring robust generalisation and avoiding artificial skill inflation due to temporal autocorrelation. </ItemContent> <ItemContent> Demonstrated that light gradient boosting machine (LGBM) outperforms other ML models, achieving a ~4% RMSE reduction and ΔR2 ≈ + 0.04 relative to the Random Forest baseline on an independent test period. </ItemContent> <ItemContent> Used SHAP-based explainable AI and independent lightning observations (WWLLN) to provide physical interpretation and establish the operational relevance of satellite-driven ML instability prediction. </ItemContent> </UnorderedList>

Satellite-driven machine learning prediction of atmospheric instability over Telangana, India: Implications for lightning activity

期刊论文
发表日期：

2026年6月22日

查看全文

Sambasivarao Velivelli,
G Ch Satyanarayana,
Hirali Sahukar,
Dodda Vyshnavi,
G Jeevana Mounika,
C V Naidu

摘要

Abstract

Accurate prediction of atmospheric instability, particularly the lifted index (LI), is critical for anticipating severe convective phenomena such as thunderstorms and lightning over tropical regions. This study develops and rigorously evaluates a satellite-driven, physically interpretable machine-learning (ML) framework for continuous regional-scale LI prediction over Telangana, India, focusing on the pre-monsoon season (April–May) using eleven years (2015–2025) of INSAT-3D/3DR Level-2 sounder observations. District-scale LI variability across ten representative districts is modelled using physically meaningful thermodynamic and dynamical predictors, including Layer-1 Precipitable Water (L1PW), wind index (WI), surface air temperature, surface pressure, and clear-sky brightness temperature at 11 μm. Six state-of-the-art regression algorithms, namely decision tree, random forest (RF), support vector regression, artificial neural network, eXtreme gradient boosting, and light gradient boosting machine (LGBM), are assessed using a temporally independent, year-wise data-partitioning strategy (training: 2015–2021; validation: 2022–2023; testing: 2024–2025) to minimize temporal autocorrelation and ensure operationally realistic generalisation. Model skill is quantified exclusively on the independent test period using multiple complementary performance metrics and a composite skill score. Results show that tree-based ensemble models consistently outperform single learners, with LGBM emerging as the most robust approach, achieving a regional-mean root mean squared error (RMSE) of 2.0 ± 0.15 K and mean coefficient of determination (R²) of 0.72, corresponding to an RMSE improvement of approximately 4% (≈ 0.08 K) and a ΔR² ≈ +0.04 relative to the RF baseline. SHapley Additive exPlanations (SHAP)-based interpretability analysis identifies L1PW and WI as the dominant physical controls on LI variability, consistent with established theories of tropical convective instability. Independent lightning observations from the World Wide Lightning Location Network (April 2024) show enhanced lightning occurrence during ML-predicted unstable LI regimes, providing an external physical consistency check rather than deterministic lightning prediction. Overall, this study demonstrates the operational feasibility of combining indigenous INSAT sounder observations with physically interpretable ML models for continuous regional instability monitoring, with clear potential for integration into IMD and TSDPS early-warning systems, offering a scalable framework for enhanced thunderstorm early warning over data-sparse regions of India.

Research highlights

Developed a satellite-driven, physically interpretable machine-learning framework for continuous Lifted Index prediction over Telangana using 11 years (2015–2025) of INSAT-3D/3DR sounder observations.

Implemented a temporally independent, year-wise validation strategy, ensuring robust generalisation and avoiding artificial skill inflation due to temporal autocorrelation.

Demonstrated that light gradient boosting machine (LGBM) outperforms other ML models, achieving a ~4% RMSE reduction and ΔR² ≈ + 0.04 relative to the Random Forest baseline on an independent test period.

Used SHAP-based explainable AI and independent lightning observations (WWLLN) to provide physical interpretation and establish the operational relevance of satellite-driven ML instability prediction.