Interpretability of compound drought-hot extreme index prediction model: a regional study in Iran

This study aims to predict a new composite drought-hot extreme index (CDHEI) that combines the standardized maximum temperature index (SMTI) and standardized precipitation index (SPI) in different climates across Iran. To this end, daily climatic data were sourced from 40 synoptic stations for the 1987-2019 period. Then, to monitor the mentioned indices simultaneously, a coupled function was employed to develop CDHEI for each climate. Three machine learning (ML) models-namely decision tree (DT), ensemble, and multilayer perceptron (MLP)-were developed to model CDHEI under three scenarios. Since machine learning models are inherently characterized by a "black box" nature, this study employed Ceteris paribus and partial dependence (CP-PD) profiles. The assessment of concurrent historical droughts and hot extremes was conducted by considering the CDHEI values and relevant categories in different climatic regions of Iran from 1998 to 2000. The results illustrated the effectiveness of the suggested index in monitoring the simultaneous occurrence of droughts and hot extreme events across different time frames and geographical areas. The most accurate ensemble model, with average values ranging from 0.1247 to 0.2047 and 0.9282 to 0.9674 (normalized root mean square error (NRMSE) and correlation coefficient (R), respectively), was the one that performed the best at the climatic zones. The CP-PD profile values demonstrated that maximum temperature had a significant effect on the model's results across all climates in scenario 2. In scenario 3, however, SPI and SMTI proved to be the most influential features.