Machine Learning Maps 37 Years of Dry Spells in India’s Cauvery Delta

In the rice heartland of Tamil Nadu, where the Cauvery River has sustained agriculture for centuries, the length of consecutive rainless days can decide whether a harvest thrives or fails. A new study published in Theoretical and Applied Climatology has now delivered the most detailed picture yet of how these dry spells behave across the Thanjavur district, combining nearly four decades of daily rainfall records with a battery of statistical tests and machine learning classifiers. The research, conducted by K. Priyanka and C. R. Suribabu of SASTRA Deemed University, analyzed daily rainfall data from 13 stations across the Cauvery Delta Zone spanning 1989 to 2025, a 37-year window long enough to separate genuine climatic signals from year-to-year noise.

Dry spells, defined as stretches of consecutive days with negligible rainfall, are among the most consequential yet underappreciated features of semi-arid hydroclimatology. Unlike droughts, which are usually assessed over months or seasons, dry spells operate on the scale of days and weeks, precisely the timescale at which crops experience water stress. For a deltaic district like Thanjavur, where paddy cultivation depends on both monsoon rainfall and canal irrigation from the Cauvery system, the duration and frequency of rainless runs determine how much supplementary water must be drawn from canals or groundwater. The researchers argue that a robust climatology of dry spell behavior is therefore a prerequisite for sustainable irrigation planning and drought vulnerability assessment in the region.

The study’s methodology unfolded in four phases. First, the team assembled a long-term climatology baseline of dry spell characteristics at each of the 13 stations. Second, they developed a descriptive, percentile-based classification scheme to sort dry spells into distinct regimes, allowing comparisons across seasons and locations. Third, they examined temporal variability and long-term trends using two widely trusted non-parametric tools: the Mann-Kendall test, which detects monotonic trends in time series without assuming any particular distribution, and Sen’s Slope estimator, which quantifies the magnitude of such trends. Finally, they turned to unsupervised and supervised machine learning to confirm that their classification reflected real structure in the data rather than arbitrary thresholds.

The machine learning confirmation relied on three complementary techniques. Principal Component Analysis (PCA) reduced the dataset to its essential dimensions, revealing that the first components captured 79.3 percent of the total variance, a strong indication that the underlying dataset has a compact and interpretable structure. K-means clustering and Agglomerative Hierarchical clustering were then applied to group stations with similar dry spell behavior. The hierarchical approach produced a dendrogram, a tree-like diagram of station similarity, that delivered one of the study’s most striking findings: the Thiruvaiyaru station stands apart from all other stations in the district, exhibiting a distinctly higher level of dry spell severity than its neighbors.

This spatial heterogeneity matters for practical water management. If dry spell risk were uniform across Thanjavur, a single district-wide irrigation schedule might suffice. Instead, the clustering results show that stations just tens of kilometers apart can belong to fundamentally different dry spell regimes, meaning that water allocation, canal scheduling, and groundwater extraction plans may need to be tailored station by station. The identification of Thiruvaiyaru as an outlier gives local authorities a concrete geographic priority for drought monitoring and for investments in storage or conveyance infrastructure that can buffer longer rainless periods.

On the temporal side, the trend analysis produced a genuinely encouraging result. Across the 37-year record, the annual dry spell duration shows a declining trend, and the regional analysis reveals that the strongest decreasing trend occurs in Thanjavur itself. In other words, over nearly four decades, the longest and most punishing rainless stretches have, on average, been shortening in this deltaic district. The authors link dry spell behavior to large-scale climate teleconnections, including the El Niño-Southern Oscillation (ENSO) and the Indian Ocean Dipole-related variability they refer to as INO, both of which are known to modulate Indian monsoon rainfall. Understanding these links opens the door to seasonal forecasting: if a developing El Niño or Indian Ocean event can be tied statistically to dry spell risk, water managers could anticipate stress months before it materializes.

The supervised learning phase of the study tested whether the dry spell classification could be predicted from hydroclimatic features alone. Two gradient- and ensemble-based algorithms, Random Forest (RF) and XGBoost, both staples of modern applied machine learning, achieved classification accuracy above 93 percent. That level of performance demonstrates that dry spell regimes in the Cauvery Delta are not stochastic noise but emerge from identifiable patterns in the hydroclimatic record. Even more valuable is the feature importance analysis: the single most powerful predictor of dry spell classification turned out to be the annual number of rainfall days. This is an intuitively satisfying result, since a landscape that receives rain on more days per year has fewer opportunities for long consecutive dry runs, but quantifying that relationship gives forecasters a simple, measurable indicator to track.

The broader scientific context makes this work timely. Across India and globally, researchers have documented shifting precipitation structures under a warming climate, with changes in the frequency and intensity of both wet and dry extremes. Studies of Indian meteorological subdivisions, the Indo-Gangetic Plains, and Mediterranean basins have all highlighted the value of classifying dry spells rather than treating rainfall as a single aggregate quantity. What distinguishes the Thanjavur study is its integration of classical climatology, rigorous non-parametric trend testing, and machine learning validation within a single framework applied at the district scale, where the results can feed directly into local decision-making rather than remaining at the level of regional generalization.

For the Cauvery Delta, where disputes over water sharing and the pressures of growing demand have made every cubic meter consequential, the study provides what the authors describe as a robust scientific basis for drought monitoring, irrigation planning, water resource management, and climate adaptation strategies. The declining trend in annual dry spell duration offers a measure of reassurance, but the pronounced spatial contrasts, exemplified by Thiruvaiyaru’s outlier status, and the demonstrated influence of ENSO and Indian Ocean variability caution against complacency. A single strong teleconnection event could still deliver a season of extended dry spells, and the percentile-based classification gives planners a common vocabulary for describing how severe such a season would be relative to the 37-year baseline.

The methodological template is arguably as important as the regional findings. By showing that PCA can compress 37 years of multi-station data into a few dominant dimensions, that clustering can expose meaningful geographic structure, and that Random Forest and XGBoost can classify dry spell regimes with over 93 percent accuracy while identifying annual rainfall days as the key predictor, the study offers a replicable workflow for any monsoon-dependent region facing similar questions. As climate variability intensifies and semi-arid agricultural zones worldwide confront longer, less predictable rainless periods, the ability to characterize, classify, and anticipate dry spells at the district scale may prove to be one of the most practical tools in the adaptation toolkit. For the farmers of Thanjavur, whose fields have depended on the rhythm of the Cauvery for generations, that rhythm has now been measured with unprecedented statistical clarity.

Subject of Research: Characterization and machine learning classification of dry spells in the Cauvery Delta district of Thanjavur, India

Article Title: The characterization of dry spell through classification and clustering analysis for cauvery river deltaic district of thanjavur: a 37-year hydroclimatic assessment

Article References: Priyanka, K., & Suribabu, C. R. (2026). The characterization of dry spell through classification and clustering analysis for cauvery river deltaic district of thanjavur: a 37-year hydroclimatic assessment. Theoretical and Applied Climatology, 157(10), Article 627. https://doi.org/10.1007/s00704-026-06554-8

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06554-8

Keywords: dry spells, Thanjavur, Cauvery Delta, drought, Mann-Kendall test, Principal Component Analysis, K-means clustering, Random Forest, XGBoost, ENSO, monsoon, irrigation planning