Leveraging external atmospheric conditions to enhance the drought predictability over the Vietnamese Mekong Delta

Severe droughts in the Vietnamese Mekong Delta (VMD) have exerted profound social and economic impacts in recent decades, underscoring the need for accurate prediction to enhance preparedness and management. Yet, the potential role of external atmospheric conditions from precipitation source regions in improving drought prediction remains underexplored. In this study, two deep learning architectures, i.e., the Convolutional Gated Recurrent Unit (ConvGRU) and Long- and Short-term Time-series Network (LSTNet), were employed to assess whether incorporating external atmospheric variables enhances drought prediction in the VMD. The ConvGRU model trained with external forcings (ConvGRU_FULL) consistently outperformed LSTNet, achieving superior skill in predicting meteorological and agricultural droughts as well as compound dry-hot events. At a 3-month lead, ConvGRU_FULL accurately identified around 90 % of meteorological and 80 % of agricultural droughts with fewer than 10 % false alarms and captured approximately 70 % and 80 % of compound dry-hot months and events, respectively. It also successfully reproduced the most severe meteorological drought and longest agricultural drought, though it underestimated the onset of the most extreme compound dry-hot event. ConvGRU_FULL predictions shifted from the overestimation to underestimation of drought severity around 2000, primarily due to data partitioning during model training, and its predictive ability at the 3-month lead showed a slight decline over time. This enhanced predictability of ConvGRU_FULL at the 3-month lead time is likely attributable to the persistent influence of external atmospheric conditions (i.e., specific humidity and wind fields) on VMD drought. Overall, ConvGRU_FULL provides a robust framework for multi-type drought prediction in the VMD.