Future drought characterization under sparse hybrid fusion weighted ensemble of global climate models

The inconsistency, complexity, and uncertainty in climate data stem from the vast proliferation of data generated by diverse climate modeling centers. Therefore, using a Multi-Model Ensemble (MME) of multiple Global Climate Models (GCMs) is essential for precisely projecting future climate. However, common statistical challenges such as non-stationarity, downscaling, model independence, model bias, spatio-temporal relative importance, redundancy, and extreme values reduce the scope of MME. This research aims to enhance the efficacy of the MME by incorporating the independent contributions of GCMs in making prediction and minimizing the influence of outliers in future drought projection. Along with the aim of minimizing computational costs, Relative Importance Metrics (RIMs) are integrated with point to point disparity in this study to address statistical challenges. Consequently, a new weighted ensemble scheme named: Sparse Hybrid Fusion Weighted Ensemble (SHFWE) and drought index named: Adaptive Multimodal Collinearity Robust Drought Index (AMCRDI) are proposed. The proposed weighting scheme assigns weights to ensemble multiple GCMs for future periods based on data from 94 grid stations of Pakistan and utilizing simulated data from 22 GCMs. The weight projection hinges on the historical coherence among simulated models and the multicollinearity among GCMs. The accuracy in AMCRDI is achieved by incorporating mixture probability models. To evaluate the efficacy of the proposed weighted ensemble, we compared its results with three competitors: Equal Weighting Averaging (EWA), Bayesian Model Averaging (BMA), and Mutual Information Weighting Scheme (MIWS). The collective numerical outcomes demonstrate that the SHFWE outperforms other methods based on error metrics, as the average error metrics for SHFWE and the competing models are (28.45, 17.96), (30.46, 19.16), (28.52, 18.37), (29.77, 19.04) in terms of Residual Mean Square Error (RMSE) and Mean Absolute Error (MAE), respectively. Subsequently, the ensemble data are employed to quantify the long-term probability of future droughts using steady-state probability analysis. The outcomes associated with steady-state probability reveal that the proposed ensemble approach exhibits a lower probability of extreme drought compared to other categories.