Adaptive ensemble learning with dynamic weighting for regional temperature extremes detection using climate teleconnections in Australia

Abstract

Australia has been labeled with an increase in temperature extremity, and this demands robust plans of adaptive detection. In this paper, an adaptive ensemble learning model is introduced that combines Support Vector Machines, Random Forest, and Gradient Boosting, and is dynamic, allowing the weighting systems to adjust over time based on performance. The Southern Annular Mode, Indian Ocean Dipole, and El Niño-Southern Oscillation are among the key drivers of the model, which is selected through correlationbased feature selection in a greedy hill-climbing approach using the ACORN-SAT data. Compared to the stationary ensembles, the weighting that drives accuracy sharpens elements of the model, thereby enhancing detection in nonstationary situations. Time cross-validation enables high cross-regime robustness. The model's accuracy is 86.4 %, its precision is 80.2 %, and its ROC-AUC is 93.4 %, which is superior to that of the single models. The driver that could be detected in feature analysis is temperature anomalies, ENSO. This solution enhances the detection of early warnings on heatwaves and addresses operational gaps in temperature monitoring.