Prépublication · 2026
Evaluation of machine-learning and deep-learning SPEI drought forecasts against persistence across timescales and lead times in Bangladesh
Mahir Ahnaf, Showmitra Kumar Sarkar — Monde
Source : arXiv
Résumé
Bangladesh, home to nearly 178 million people, is highly vulnerable to recurring droughts that threaten food security and rural livelihoods. Machine-learning and deep-learning models are increasingly used to forecast drought indices such as the Standardized Precipitation Evapotranspiration Index (SPEI). Most studies report accuracy without a persistence reference, although accumulated indices are strongly autocorrelated, and seldom test drought detection, longer leads or variation between training runs. This study forecast SPEI at 3-, 6- and 12-month timescales one to three months ahead for th…