A spatial duration-augmented framework for drought persistence
Drought monitoring across heterogeneous climate regions requires explicit modeling of drought persistence rather than relying solely on index classification. We develop a Bayesian spatial, duration-augmented framework for drought persistence that captures both temporal memory and spatial dependence. Because persistence estimates depend on the underlying drought index, we also assess the influence of (i) reference evapotranspiration (RET) formulations and (ii) the choice of probability distribution used to standardize the climatic water balance. The framework is applied to a high-resolution (0.25$^\circ \times$ 0.25$^\circ$) dataset over 730 grid cells across Italy. The choice of RET formulation (Penman-Monteith, Hargraves, Thornthwaite) has only a marginal influence on the overall drought signal: spatially averaged SPEI series at the 3-, 6-, and 12-month scales are identical across methods. By contrast, the proposed bulk-and-tails (BATs) distribution substantially improves index estimation, consistently outperforming commonly used alternative models across all RET methods and timescales and achieving normality acceptance rates of 98.4%--99.9% compared to 37%--92% for the generalized extreme value (GEV) distribution and 0%--30% for others. The spatial duration-augmented model further outperforms a nonspatial counterpart, producing spatially continuous, uncertainty-aware recovery and survival fields. Results reveal a strong south-north gradient in temporal memory: southern peninsular and island regions (Sicily, Sardinia) exhibit pronounced persistence and duration-dependent recovery hazards, whereas northern and Alpine regions show near-zero duration dependence. Drought characterization is more sensitive to the distributional and persistence assumptions than to the RET formulation, and duration-insensitive classifications systematically understate extended drought risk in southern Italy.