Abstract Accurate assessment and future characterisation of regional drought are critical for formulating effective mitigation policies. Global climate models (GCMs) are widely used to evaluate future drought conditions; however, selecting the most relevant models for a specific region remains a challenge. Among various machine learning techniques, the Boruta algorithm, based on the random forest (RF) classifier, offers significant advantages in feature selection and ranking, as it identifies all relevant predictors rather than just a minimal subset. In this study, 22 GCMs were evaluated using the Boruta algorithm to identify the most significant models for subsequent drought assessment. The top-performing models, BCC-CSM2-MR, CNRM-CM6-1, CNRM-CM6-1-HR, CNRM-ESM2-1, and IPSL-CM6A-LRac, achieved net strength scores of 216, 180, 150, 166, and 166, respectively, with no associated weakness scores. In contrast, models such as GFDL-ESM4, with a net strength of –181, demonstrated the poorest performance. To enhance regional drought characterisation, a novel framework called the multimodel Boruta standardised drought index (MBSDI) was developed and applied to long-term time series data from 22 GCMs across 21 locations in the Sindh Province of Pakistan. The historical period considered spans from 1950 to 2014, while future drought projections are evaluated under three shared socioeconomic pathways (SSPs), SSP1-2.6, SSP2-4.5, and SSP5-8.5, for the period from 2015 to 2100. For future characterisation, steady-state probabilities derived from Markov chains were used to quantify the persistence and severity of drought conditions over time. The results strongly support the effectiveness of the MBSDI in regional drought assessment, particularly under varying climate scenarios. Projections indicate increasing drought severity under high-emission pathways, especially over extended time scales. Overall, this study presents a robust statistical framework for regional drought analysis using multiple GCMs and offers valuable insights to support evidence-based decision-making for drought mitigation and climate adaptation strategies. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Boruta-based GCM selection ensuring robustness and unbiased selection of high-performing GCMs for drought assessment. </ItemContent> <ItemContent> Development of a novel drought index named MBSDI, which integrates multiple GCM weights into a unified probabilistic drought index, enhancing regional drought characterization across spatial and temporal scales. </ItemContent> <ItemContent> Markov chain–based steady-state probability analysis effectively quantifies drought persistence and long-term severity under multiple climate scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5). </ItemContent> <ItemContent> Future projections reveal intensified drought conditions, particularly under high-emission scenarios, providing critical insights for climate adaptation and drought mitigation planning. </ItemContent> </UnorderedList>

A hybrid machine learning–Markov chain framework for GCM subset selection and probabilistic regional drought assessment

期刊论文
发表日期：

2026年7月6日

查看全文

Muhammad Shakeel,
Zulfiqar Ali,
Veysi Kartal,
Hussnain Abbas,
Maysaa Elmahi Abd Elwahab,
Rizwan Niaz

摘要

Abstract

Accurate assessment and future characterisation of regional drought are critical for formulating effective mitigation policies. Global climate models (GCMs) are widely used to evaluate future drought conditions; however, selecting the most relevant models for a specific region remains a challenge. Among various machine learning techniques, the Boruta algorithm, based on the random forest (RF) classifier, offers significant advantages in feature selection and ranking, as it identifies all relevant predictors rather than just a minimal subset. In this study, 22 GCMs were evaluated using the Boruta algorithm to identify the most significant models for subsequent drought assessment. The top-performing models, BCC-CSM2-MR, CNRM-CM6-1, CNRM-CM6-1-HR, CNRM-ESM2-1, and IPSL-CM6A-LRac, achieved net strength scores of 216, 180, 150, 166, and 166, respectively, with no associated weakness scores. In contrast, models such as GFDL-ESM4, with a net strength of –181, demonstrated the poorest performance. To enhance regional drought characterisation, a novel framework called the multimodel Boruta standardised drought index (MBSDI) was developed and applied to long-term time series data from 22 GCMs across 21 locations in the Sindh Province of Pakistan. The historical period considered spans from 1950 to 2014, while future drought projections are evaluated under three shared socioeconomic pathways (SSPs), SSP1-2.6, SSP2-4.5, and SSP5-8.5, for the period from 2015 to 2100. For future characterisation, steady-state probabilities derived from Markov chains were used to quantify the persistence and severity of drought conditions over time. The results strongly support the effectiveness of the MBSDI in regional drought assessment, particularly under varying climate scenarios. Projections indicate increasing drought severity under high-emission pathways, especially over extended time scales. Overall, this study presents a robust statistical framework for regional drought analysis using multiple GCMs and offers valuable insights to support evidence-based decision-making for drought mitigation and climate adaptation strategies.

Research highlights

Boruta-based GCM selection ensuring robustness and unbiased selection of high-performing GCMs for drought assessment.

Development of a novel drought index named MBSDI, which integrates multiple GCM weights into a unified probabilistic drought index, enhancing regional drought characterization across spatial and temporal scales.

Markov chain–based steady-state probability analysis effectively quantifies drought persistence and long-term severity under multiple climate scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5).

Future projections reveal intensified drought conditions, particularly under high-emission scenarios, providing critical insights for climate adaptation and drought mitigation planning.