Machine learning taps ocean-climate signals to sharpen India’s drought forecasts

Seasonal forecasts of rain and temperature over India have long been hampered by a stubborn problem: the dynamical models that drive them carry errors too nonlinear and too tangled with large-scale climate oscillations for conventional statistical correction to fix. A new study published in Climate Dynamics proposes a way through that bottleneck. Researchers Paushali Deb, Saurabh Verma and L. Karthikeyan built a calibration framework that feeds machine learning algorithms not only the raw output of seasonal forecast models, but also a curated set of teleconnection drivers — the Madden-Julian Oscillation at four longitudes (80°, 100°, 120° and 140°E), the Indian Ocean Dipole, the North Atlantic Oscillation, the Pacific Decadal Oscillation and the Niño 3.4 index. The idea is that the same large-scale signals that make Indian rainfall so hard to predict also carry information that, if handled properly, can make the forecasts better.

The team calibrated four ensemble forecast systems spanning the North American Multi-Model Ensemble and ECMWF archives — GEOSS2S, CFSv2, CCSM4 and SEAS5 — at lead times of two and three months. Two tree-based machine learning methods, Random Forest and XGBoost, did the calibrating, and their performance was benchmarked against two established statistical post-processing techniques: quantile mapping and canonical correlation analysis. The evaluation rested on a 42-year record with strict temporal separation, meaning the algorithms learned only from data between 1982 and 2016 and were then judged on entirely independent forecasts from 2017 to 2023. That design matters, because post-processing methods can look impressive when tested on the data they were trained on, and the authors deliberately closed that door.

Across variables, lead times and regions, Random Forest delivered the most consistent gains. The researchers attribute this to its bagging-based ensemble structure, which reduces forecast variance and, crucially, captures the nonlinear interactions between teleconnections and precipitation that linear statistical methods cannot represent. The machine learning approach also substantially narrowed the skill gap between two-month and three-month lead forecasts for both precipitation and temperature — a persistent weakness of seasonal prediction, where skill typically decays rapidly as the forecast horizon extends. Beyond raw forecast variables, the framework improved meteorological drought detection skill over agriculturally sensitive regions of India, which is where the practical stakes are highest for a country whose farming economy remains deeply exposed to monsoon variability.

One of the more revealing parts of the study is its use of SHAP-based feature attribution, a technique that traces which predictors the machine learning models actually relied on. At shorter lead times, the North Atlantic Oscillation and the Madden-Julian Oscillation at 120°E dominated the calibration skill. At the three-month lead, more persistent large-scale modes gained relative importance — a pattern the authors interpret as consistent with the decay of short-lived signals at extended forecast horizons. In other words, the machine learning models were not just memorising correlations; the shifting hierarchy of predictors they learned mirrors the physical lifetimes of the climate modes themselves, which lends the framework a degree of interpretability that black-box forecasting tools often lack.

The broader significance lies in what the authors describe as a physically informed pathway for forecast improvement. Rather than treating calibration as a purely statistical exercise, the framework explicitly encodes the climate drivers that shape Indian rainfall and temperature, allowing the algorithms to exploit relationships that conventional post-processing leaves on the table. The findings suggest that teleconnection-aware machine learning calibration provides a robust and interpretable route to sharper seasonal drought outlooks across the climatologically diverse regions of India, in a forecasting landscape where uncertainties in dynamical models substantially degrade predictive skill.

Subject of Research: Teleconnection-aware machine learning calibration of multi-model seasonal forecasts for drought prediction over India

Article Title: Teleconnection-aware machine learning calibration of multi-model seasonal forecasts for drought prediction over India

Article References: Deb, P., Verma, S., & Karthikeyan, L. (2026). Teleconnection-aware machine learning calibration of multi-model seasonal forecasts for drought prediction over India. Climate Dynamics, 64(11), Article 480. https://doi.org/10.1007/s00382-026-08419-2

Image Credits: AI Generated

DOI: 10.1007/s00382-026-08419-2

Keywords: seasonal forecasting, machine learning, drought prediction, India, teleconnections, Madden-Julian Oscillation, North Atlantic Oscillation, Indian Ocean Dipole, Random Forest, XGBoost, monsoon, Climate Dynamics