A novel spatially coherent graph-cut–optimized multi-model ensemble framework with ML weighting for advanced drought assessment under climate change

Abstract

ABSTRACT Reliable drought assessment requires climate projections that capture spatial and temporal precipitation variability. Global Climate Models (GCMs) often exhibit structural biases and inconsistent regional performance. Traditional station-wise model selection also produces fragmented and physically implausible spatial patterns. This study proposed Spatio-Temporal Adaptive GCM Framework for Multi-Model Ensemble and Drought Assessment (STAG-MDA), applied to Pakistan using monthly precipitation data from 1950 to 2009 across 94 grid stations. STAG-MDA combines decadal model evaluation with a spatially coherent graph cut (GC) optimization, minimizing an energy function balancing data fidelity and spatial smoothness to identify the most suitable GCM per grid cell. Selected models are further integrated using a five-strategy machine-learning ensemble system with adaptive, performance-based weights. The XGBoost-weighted ensemble (XGBoostE) delivered the strongest improvements, reducing MAE by 18–27%, lowering normalized NRMSE by 15–22%, and increasing Kling Gupta efficiency (KGE) by 20–35%, with 92 regional wins. Gaussian mixture modeling captures multimodal and heavy-tailed precipitation behavior, supporting the development of the STASPI-GM drought index, which improves characterization of drought persistence, severity, and evolution across timescales. Scenario-based projections under SSP1-2.6, SSP2-4.5, and SSP5-8.5 indicate increasing drought severity and frequency. Although focused on drought, STAG-MDA provides framework for GCM evaluation and uncertainty reduction.