A Novel Multi-Model Climate Ensemble Approach to Assess Hydrological Drought Under Climate Change With Reduced Uncertainty

Abstract

Drought is a natural outcome of global warming which brings serious effects in various fields. To avert its consequences and facilitate long-term planning, proper forecasting of droughts is critical. Global climate models (GCMs) are useful tools to forecast hazards related to climate. But GCMs are biased, uncertain, and interdependent. Multi-model ensembles (MMEs) are useful to overcome these deficiencies. However, even the classical ensemble methods like equal model averaging (EMA) and mutual information weighting still have the weaknesses of being sensitive to outliers and less robust in highly variable areas. This study intended to propose a new machine learning-based solution, the K-Nearest-Neighbors Adaptive Weighting Scheme (KNNAWS-Ensemble) that resolves the discrepancies in traditional weighting approaches and enhances the accuracy of the drought forecasting. KNNAWS-Ensemble uses the K-Nearest Neighbors (KNN) algorithm to assign adaptive weights to 18 CMIP6 GCMs that are based on historical performance. As a reference data set, precipitation data of the Tibetan Plateau is utilized from 1961 to 2014. Additionally, K-Component Gaussian Mixture Model (K-CGMM) is used for the statistical distribution fitting and standardization. KNNAWS-Ensemble is validated across conventional ensemble methods using quality measures such as the normalized root mean square error (NRMSE), mean absolute error (MAE), and the correlation coefficient. The findings prove that the KNNAWS-Ensemble is more effective than the conventional ensemble methods. Meanwhile, it has the lowest NRMSE (7.2560) and MAE (1.4127) and largest correlation coefficient (0.9074). Conversely, there are greater values of NRMSE in EMA and the Mutual Information Ensemble (MIE) (8.5516 and 8.4122), and lower correlation values (0.8958 and 0.8960). These results verify high predictive validity and less uncertainty of the proposed scheme. Furthermore, long-term drought projections underscore the elevated risk of severe drought (SD) events. Under SSP1-2.6, the average SD probability increased from 0.533 at the 1-month scale to 0.705 at the 48-month scale. Similarly, under SSP2-4.5, average SD exhibits persistently high probabilities, ranging from 0.668 (1 month) to 0.681 (48 months), while under the high-emission SSP5-8.5 scenario, the average probability SD is 0.406 at the 48-month scale. These findings reveal the higher probabilities of severe drought than other categories under future climate scenarios. The KNNAWS-Ensemble offers an accurate framework on prediction of drought and is useful to address uncertainties and outlier sensitivity in MMEs. Such development serves as a good aid to managing and mitigating droughts especially in case of rising climatic variability.