Spatial clustering of drought and wet spell regimes over Italy: a feature-based network approach across multiple SPEI timescales

This study introduces a network-based framework for characterizing and clustering the spatial organization of drought and wet extreme regimes at multiple accumulation timescales, applied to the 0.5° gridded SPEI Global Drought Monitor dataset over Italy. Extreme events are identified through run theory, extracting a set of statistical descriptors for each pixel and temporal scale: total number and frequency of events, temporal event occurrence fraction, mean and maximum duration, mean and maximum severity, mean and maximum intensity, global and local inter-event temporal irregularity, and the dominant season of event peak occurrence. These descriptors constitute a 12-dimensional feature space, which is transformed into a similarity network using Mahalanobis distance to account for inter-feature correlations. Community detection is then performed via the Louvain algorithm to identify pixel clusters sharing statistically similar extreme regimes. Partition quality is assessed through modularity Q and feature-space silhouette width and validated against a 1,000-sample permutation-based null distribution.

The Louvain algorithm consistently identifies 4 to 6 communities across all temporal scales for both drought and wet events, with all partitions achieving high statistical significance. For drought events, modularity Q exhibits a clear increasing trend at longer accumulation periods, rising monotonically from SPEI-7 to SPEI-12, indicating that community structure becomes progressively more pronounced as the accumulation timescale increases. F-ratio analysis reveals a clear three-phase regime across timescales. At short scales (SPEI-1, SPEI-3, SPEI-5), the dominant season of event peak occurrence governs cluster discrimination, reflecting the strong seasonality of short-term drought events. At intermediate scales (SPEI-6 to SPEI-9), discrimination shifts toward event-structure metrics such as severity and frequency. At long scales (SPEI-10 to SPEI-12), maximum spell severity and duration reach their highest explanatory power, indicating that persistent multi-month drought structures become the dominant axis of spatial differentiation.

For wet events, modularity Q does not exhibit a monotonic trend across scales and remains lower than its drought counterpart at most timescales, suggesting that wet regimes are spatially less structured than drought regimes. The dominant season of event peak occurrence is nearly uninformative for clustering, indicating an absence of seasonally organized differentiation. From SPEI-1 to SPEI-9, cluster discrimination is primarily driven by maximum duration and severity, peaking at SPEI-4 and SPEI-5. A sharp regime shift occurs at longer timescales (SPEI-10 and SPEI-11), where these duration- and severity-based features lose most of their discriminative power, and are replaced by mean intensity and inter-event temporal irregularity, revealing a scale-dependent reorganization of the drivers of spatial heterogeneity in wet spell regimes.

This study, integrating run theory, multi-scale feature extraction, and network-based community detection, provides a statistically robust and transferable framework for characterizing the spatial organization of hydroclimatic extremes, whose scale-dependent and phase-specific clustering structure would remain hidden to conventional single-scale approaches.