AI-Driven Lightning Risk Mapping in Bihar, India: An Ensemble Machine Learning Approach

Abstract

Lightning strikes are a recurrent, spatially concentrated hazard in Bihar, India, with peak incidence between May and September driving substantial casualties and economic losses. This study presents a district-level lightning vulnerability assessment for all 38 districts of Bihar (2019–2025), integrating geospatial analysis with an ensemble machine learning (ML) framework combining Random Forest (RF), linear regression (LR), support vector regression (SVR), and artificial neural networks (ANN). Feature selection using Pearson correlation (p < 0.05) and Variance Inflation Factor screening (VIF < 10) retained four predictors—mean elevation, population density, fallow land, and scrub land—with fallow land (r = 0.76), mean elevation (r = 0.70), and net CG strokes (r = 0.78) showing the strongest associations with casualties. RF feature importance ranked mean elevation (62.5%), fallow land (10.6%), and population density (10.1%) as dominant contributors. Model validation via Leave-One-Out Cross‑Validation (LOOCV, N = 38), supplemented by spatially aware Leave-One-Division-Out (LODO) block cross-validation across Bihar's nine divisions, identified SVR as the strongest individual model (R 2 = 0.666, RMSE = 3.21), followed by LR (R 2 = 0.625), ANN (R 2 = 0.550), and RF (R 2 = 0.515). An inverse-RMSE-weighted ensemble achieved the highest accuracy (R 2 = 0.668, RMSE = 3.20, MAE = 2.49), with quintile-stratified analysis confirming an 8.7% RMSE reduction over LR at the high-casualty tail. The narrow LOOCV–LODO gap (ΔR 2 = +0.030) indicates genuine model complementarity rather than spatial overfitting. A district-level vulnerability index identified Gaya, Aurangabad, Rohtas, Jamui, Banka, and West Champaran as highest-risk, with spatial patterns mirroring CG stroke frequency, supporting early-warning systems, awareness campaigns, and infrastructure planning to reduce lightning casualties across Bihar.