Interpretable Unsupervised Identification of Persistent Rainfall Regimes with Robustness and Temporal Stability Assessment

Resumen

Rainfall variability strongly constrains agricultural planning, drought-risk management, and climate resilience. Yet many large-scale rainfall regime studies provide limited evidence on robustness across clustering methods and on temporal persistence of the discovered regimes. This study proposes an interpretable unsupervised artificial intelligence framework to identify dominant rainfall regimes across 36 Indian meteorological subdivisions using monthly rainfall records spanning 1901–2015. We engineered transparent climatic descriptors capturing rainfall magnitude, interannual variability, seasonal peak intensity, and intra-annual concentration. K-Means clustering with silhouette-based model selection favored a two-regime partition (k = 2), achieving a Silhouette coefficient of 0.428, indicating moderate but meaningful separability. Robustness was evaluated by applying Ward hierarchical clustering to the same standardized feature space, yielding comparable separation (Silhouette = 0.413) and strong cross-method agreement (Adjusted Rand Index, ARI = 0.881). Statistical separability was confirmed using one-way ANOVA with large effect sizes (η² up to 0.860 for mean annual rainfall), highlighting annual magnitude and peak intensity as primary discriminative drivers. Temporal replication across three historical sub-periods (1901–1950, 1951–1980, 1981–2015) showed high persistence: 35 of 36 subdivisions retained stable regime membership, with only Lakshadweep transitioning once (P1→P2). The resulting pipeline is reproducible, physically interpretable, and directly relevant for climate-risk zoning and long-term water-resource and agricultural planning.