Multi-Index Drought Projections Under CMIP6 Scenarios in Northwestern Bangladesh: A Dual-Scale Model Screening and Explainable Machine Learning Approach

Rajshahi Division, sitting on Bangladesh's semi-arid Barind Tract, has seen drought happenings clearly intensify lately, putting more stress on farming and groundwater resources. Earlier studies trace drought patterns using standardized indices and climate models, but quite a few lean too much on just one measure, choose models without the right checks, and stop before explaining the actual physical steps that let drought intensify.

To get around those weak points, we push thirteen bias-corrected CMIP6 models through a workflow from model screening to drought computation and then perform driver analysis. We test how models match the Bangladesh Meteorological Department's logs and ERA5 reanalysis at two levels, the closest grid cell and a regional spatial average, because validating at only one scale was insufficient. We then rank the models using a TOPSIS routine based on KGE scores, RMSE, and seasonal rainfall, applying bootstrap resampling to validate the outcomes.

Next, we analyze three drought metrics: SPI, SPEI, and RDI. We also use the Hargreaves–Samani with Penman–Monteith equations, just to see how the PET method tweaks expected drought intensity. Simultaneously, we quantify ETCCDI climate extremes, and then add a compound hot–dry score, describing the co-occurrence of extreme heat alongside precipitation shortage across four Shared Socioeconomic Pathways.

After that, we train an XGBoost model and use SHAP to sort out which seasonal climate variables raise or lower the number of droughts. This is meant to make the drivers more legible, determining what exactly is pushing the droughts harder.

So far, the preliminary findings point to a pre-monsoon evapotranspiration spike and compound hot–dry days as key engines of the projected drought severity, especially under SSP3-7.0 and SSP5-8.5. Also, the difference between Penman–Monteith and Hargreaves-based results increases a lot in later decades, demonstrating that the drought outlook is very sensitive to how PET is formulated. Overall, these outputs matter for designing early warning systems in areas that do not have sufficient ground monitoring infrastructure.