To further understand the model performance, we have performed additional model predictions for severe drought events with SPEI below −1 (Figure 8 and Table 3). In comparison to Figure 5 and Table 2, the model performance is low. Except for XGB, the AUCs are well below 1. It may be attributed to the fact that the sample size of drought impact on public water supply observed with SPEI less than −1 is not large enough for stable prediction (refer to Figure 3). Despite the deficient sample size, these results demonstrate the advantages of the XGB, with a value of 0.97 for the nationwide case. Note that the decisions for spatial sampling may affect the model performance. SPEI is a standardized variable and SPEI determines drought conditions with the same occurrence frequency everywhere (Blauhut et al. 2016). And the use of the mean SPEI, and the grouping of the samples into regional scales limits the precision of the classification between times with or without impacts and the corresponding drought indices. We expect as data availability increases, the analyses could be repeated at smaller spatial units.