Running summarizing tools on a new article

Check the paper below, then start its synthesis.


Evaluation of machine-learning and deep-learning SPEI drought forecasts against persistence across timescales and lead times in Bangladesh

Mahir Ahnaf, Showmitra Kumar Sarkar

arXiv: 2611.25075v1 (physics.ao-ph)
95 pages (manuscript 55, Supplementary Information 40), 15 figures, 4 tables; Supplementary Information included (5 figures, 18 tables). Code: https://github.com/mahirahnaf1727/bangladesh-spei-drought-forecasting (archived at https://doi.org/10.5281/zenodo.24120523). Data: https://doi.org/10.5281/zenodo.24115493
License: CC BY 4.0

Abstract: 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 the eight divisions of Bangladesh, using ERA5 reanalysis checked against 35 Bangladesh Meteorological Department stations. Five models, XGBoost, CNN, Bi-LSTM, CNN-BiLSTM and CEEMDAN-BiLSTM, were trained with five seeds each, scored on accuracy, skill against persistence and drought detection, and interpreted with Shapley values. At a one-month lead, the CNN-BiLSTM and Bi-LSTM beat persistence clearly at SPEI-3 and SPEI-6 (skill 0.408 and 0.350), while at SPEI-12 a Nash-Sutcliffe efficiency of 0.872 barely exceeded persistence (0.850). Skill and detection could disagree: at SPEI-3 and a three-month lead, XGBoost had the highest skill (0.439) yet detected none of the 119 drought months. Only at SPEI-6 did the forecasts beat persistence significantly and match its detection at all three leads, with fewer false alarms. At SPEI-12, gains were small or absent and varied between seeds. The forecasts kept 76-80% of their correlation against station-derived SPEI and relied mainly on precipitation and lagged SPEI. Forecasts should therefore be judged on skill against persistence and detection across several seeds, not on accuracy alone. In Bangladesh, SPEI-6 forecasts can support early warning up to three months ahead and SPEI-3 forecasts one month ahead, while SPEI-12 is better suited to monitoring.

Submitted to arXiv on 18 Nov. 2026