Cutting-edge trends in drought prediction and mitigation using ensemble learning and augmented XAI: a meteorological study
Abstract
Drought affects more than 60% of India's population, reducing agricultural productivity, straining water resources, and threatening socioeconomic stability. This study proposes an interpretable one-month-ahead meteorological drought prediction framework by integrating class-weighted ensemble machine learning with SHAP and LIME explainable artificial intelligence (XAI). Historical meteorological observations (1972–2022) obtained from the Zenodo climate database were analyzed for 49 districts across five drought-prone Indian states (Bihar, Maharashtra, Gujarat, Andhra Pradesh, and Karnataka). After generating lagged meteorological features (t − 1 to t − 6) and addressing severe class imbalance (7.07% drought events), multiple machine learning classifiers were evaluated. The class-weighted XGBoost model (scale_pos_weight ≈ 11.90, without SMOTE) achieved the best overall performance on the independent 2016–2022 test dataset, with an accuracy of 95.7%, macro precision of 84.7%, macro recall of 81.0%, macro F1-score of 82.7%, and a PR-AUC of 0.647 on the highly imbalanced test set (outperforming both unweighted training and SMOTE-based approaches). SHAP and LIME consistently identified lagged Standardized Precipitation Index (SPI), lagged precipitation, vapour pressure, and wet-day frequency as the most influential drivers of drought occurrence. Exploratory temporal analysis further revealed declining wet-day frequency in Karnataka, increasing vapour pressure in Andhra Pradesh, and substantial climatic variability across the study regions. The proposed framework combines accurate drought prediction with transparent model interpretation, providing a practical decision-support tool for district-level drought early warning, water resource management, and climate-resilient agricultural planning in India.
Related
- Ensemble fire probability prediction by integrating meteorological AI models with lightweight prediction networks
- Ensemble tree-based machine learning models for data-driven AQI prediction
- Quantifying meteorological impacts on local landfill methane emissions by using field measurements and machine learning
- Sentinel-1 SAR time series for forest phenology: multimodal deep learning with meteorological fusion across deciduous broadleaf and evergreen needleleaf forests
- Climate-stratified assessment of PM2.5 using machine learning: Geographic controls dominate meteorological factors
BibTeX
@article{rathi2026cuttingedge,
title = {{Cutting-edge trends in drought prediction and mitigation using ensemble learning and augmented XAI: a meteorological study}},
author = {Megha Rathi and Shailendra Mishra and Naif S. Alshammari and Himani Agrawal and Ehtesham Ashraf},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-69632-x},
url = {https://www.nature.com/articles/s41598-026-69632-x},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Cutting-edge trends in drought prediction and mitigation using ensemble learning and augmented XAI: a meteorological study
AU - Megha Rathi
AU - Shailendra Mishra
AU - Naif S. Alshammari
AU - Himani Agrawal
AU - Ehtesham Ashraf
PY - 2026
DA - 2026/10/22
JO - Scientific Reports
DO - 10.1038/s41598-026-69632-x
UR - https://www.nature.com/articles/s41598-026-69632-x
AB - Drought affects more than 60% of India's population, reducing agricultural productivity, straining water resources, and threatening socioeconomic stability. This study proposes an interpretable one-month-ahead meteorological drought prediction framework by integrating class-weighted ensemble machine learning with SHAP and LIME explainable artificial intelligence (XAI). Historical meteorological observations (1972–2022) obtained from the Zenodo climate database were analyzed for 49 districts across five drought-prone Indian states (Bihar, Maharashtra, Gujarat, Andhra Pradesh, and Karnataka). After generating lagged meteorological features (t − 1 to t − 6) and addressing severe class imbalance (7.07% drought events), multiple machine learning classifiers were evaluated. The class-weighted XGBoost model (scale_pos_weight ≈ 11.90, without SMOTE) achieved the best overall performance on the independent 2016–2022 test dataset, with an accuracy of 95.7%, macro precision of 84.7%, macro recall of 81.0%, macro F1-score of 82.7%, and a PR-AUC of 0.647 on the highly imbalanced test set (outperforming both unweighted training and SMOTE-based approaches). SHAP and LIME consistently identified lagged Standardized Precipitation Index (SPI), lagged precipitation, vapour pressure, and wet-day frequency as the most influential drivers of drought occurrence. Exploratory temporal analysis further revealed declining wet-day frequency in Karnataka, increasing vapour pressure in Andhra Pradesh, and substantial climatic variability across the study regions. The proposed framework combines accurate drought prediction with transparent model interpretation, providing a practical decision-support tool for district-level drought early warning, water resource management, and climate-resilient agricultural planning in India.
SN - 2045-2322
ER -