Uncertainty-aware deep learning for national lightning hazard and risk mapping to support sustainable disaster risk reduction in Bangladesh

Fardin Bin Faruque , Kazi Tahmid Parveg, Jannatul Ferdous

Bangladesh has one of the world's highest lightning death tolls, borne mostly by farmers and fishers, yet national hazard maps rest on interpolation or tabular machine learning without independent validation. We developed a raster-native deep-learning framework reproducing monthly lightning stroke density from the WWLLN Global Lightning Climatology (WGLC, 2012-2025) on a 5-arc-minute grid from 18 meteorological, aerosol and surface predictors, comparing U-Net, Attention U-Net, ConvLSTM and a ConvLSTM-U-Net hybrid under a predefined protocol. On unseen years (2023-2025), every architecture outperformed monthly climatology, and a 12-member ensemble cut the mean squared error by 20 % and raised the hotspot critical success index from 0.15 to 0.27, whereas a convective available potential energy (CAPE) × precipitation proxy failed on hotspots (0.08). Recalibrated ensemble spread gave 90 % coverage of 95 % intervals. The ensemble matched the independent TRMM-LIS/ISS-LIS satellite climatology (r = 0.53) about as well as the ground record did, and Integrated Gradients ranked CAPE first in all 12 runs. A change-point test found a 2016 detection step (×1.96) in the observed record absent from the meteorological drivers, so no trend is claimed. With hazard, exposure and vulnerability combined under propagated uncertainty, risk peaks in the north-eastern haor basin, where 50 upazilas (15.5 million people) have probability ≥ 0.9 of being high or very high risk. District deaths follow pre-monsoon risk (ρ = 0.40) more closely than annual risk (ρ = 0.19), supporting harvest-timed warnings and shelters targeted to these upazilas under Sustainable Development Goal targets 1.5, 11.5 and 13.1.