Laghari et al. (2025) Predicting spatiotemporal changes in flood prone regions using PSO-ML coupling under climate change scenarios

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-02
• Authors: Azhar Ali Laghari, Yongheng Shen, Akash Kumar, Qurat-ul-ain Abro, Yanli Shen, Qingxia Guo
• DOI: 10.1038/s41598-025-26939-5

Research Groups

• College of Resources and Environment, Shanxi Agricultural University, Jinzhong, Shanxi Province, China
• School of Civil Engineering, Guangzhou University, Guangzhou, China
• Business School, University of International Business and Economics, Beijing, China

Short Summary

This study developed Particle Swarm Optimization-Machine Learning (PSO-ML) models, integrated with General Circulation Model (GCM) data, to predict spatiotemporal flood risk in Shanxi Province, China, under climate change scenarios. The PSO-ML models significantly improved prediction accuracy, projecting a southward shift and increase in flood-prone areas by 2100, with land use, elevation, and slope being the most influential factors.

Objective

• To develop and apply Particle Swarm Optimization-Machine Learning (PSO-ML) models, integrated with General Circulation Model (GCM) data, to explore and predict the spatio-temporal distribution of flood disasters in Shanxi Province, China, under various climate change scenarios.
• To address the precision shortcomings of traditional regional flood disaster risk assessment methods.

Study Configuration

• Spatial Scale: Shanxi Province, China (110°14′–114°33′ E, 34°34′–40°44′ N), with a total area of 1.56 × 10^6 km². Data resolutions varied from 0.01° to 1 km, with GCM data resampled to 0.1° × 1°. Spatial statistical unit was county level.
• Temporal Scale: Historical data covered 1981–2023 (extreme precipitation) and 2016–2023 (historical flood inundation areas). Climate model historical simulations spanned 1860–2014, and future predictions were made for 2030, 2050, 2070, and 2100 based on GCM future projections from 2015–2100.

Methodology and Data

• Models used:
  • Machine Learning (ML) models: Random Forest (RF), Support Vector Classification (SVC), XGBoost, AdaBoost, K-Nearest Neighbors (KNN), Graph Neural Networks (GNN).
  • Optimization algorithm: Particle Swarm Optimization (PSO).
  • Coupled models: PSO-RF, PSO-SVC, PSO-XGBoost, PSO-AdaBoost, PSO-KNN, PSO-GNN.
  • Climate Model: Centre National de Recherche Météorologique Climate Model version 6.1 - High Resolution (CNRM-CM6-1-HR) from CMIP6.
  • Bias correction method: Delta (Δ) bias correction.
• Data sources:
  • Precipitation data: Pentad high-resolution precipitation dataset (Google Earth Engine, 1981-2023, daily, county level); Spatiotemporal Tri-pole Environment big data platform (monthly average, 1 km resolution); ECMWF Reanalysis v5 (ERA5) (0.1° resolution, corrective data); CNRM-CM6-1-HR (CMIP6, 0.5° × 0.5° raw, resampled to 0.1° × 1°).
  • Flood historical data: Sentinel-1 A SAR (Synthetic Aperture Radar) data (Google Earth Engine platform), USGS Earth Explorer website (0.01° resolution), covering 2016-2023.
  • Flood condition factors (15 total):
    • Elevation (DEM): USGS Earth Explorer website (0.01°).
    • Derived topographic factors: Slope, Aspect, Curvature, Topographic Wetness Index (TWI).
    • Meteorological/Hydrological: Precipitation (Pr), Temperatures (Tas), Soil water storage (Mrsos), Surface runoff (Mrros) (from Earth System Grid Federation (ESGF) and ERA5/CMIP6).
    • Geographical/Anthropogenic: Geological type (Natural Resources Canada, 0.01°), Soil type (A Big Earth Data Platform for Three Poles, 0.01°), Land Use/Cover Change (LUCC) (Scientific Data (nature.com), 0.05°), Distance to River, Distance to Railway, Distance to Highway.

Main Results

• The frequency of days with daily precipitation exceeding 50 mm in Shanxi Province increased significantly from 23 in 1981 to 71 in 2021. Spatially, this trend showed a gradual increase from north (Datong, 18 days) to south (Yuncheng, 71 days).
• PSO-ML models demonstrated superior performance compared to traditional single ML models. PSO-XGBoost, PSO-RF, and PSO-KNN achieved the highest training AUC values of 0.98, 0.95, and 0.94, respectively. Validation AUCs were 0.94 for PSO-XGBoost, 0.91 for PSO-KNN, and 0.89 for PSO-RF.
• The most influential factors for flood susceptibility were identified as land use change (LUCC) (10.37%), elevation (DEM) (10.01%), and slope (8.76%), followed by soil water storage (Mrsos) (8.05%) and distance to road (7.72%).
• Under four Shared Socioeconomic Pathways (SSP126, SSP245, SSP370, SSP585), flood-prone areas in Shanxi Province are projected to shift southward.
• The SSP370 scenario projects the slowest growth, reaching 7660.116 km² of at-risk area by 2100.
• The SSP585 scenario projects the most rapid growth, peaking at 13,933.69 km² by 2070, before a slight decrease.
• Overall, flood-prone plots are projected to increase from 3.15% (before 2021) to between 5.71% (SSP370) and 11.10% (SSP245) by 2100.

Contributions

• Proposes a novel and highly efficient Particle Swarm Optimization-Machine Learning (PSO-ML) ensemble modeling system for flood disaster risk assessment, integrating multiple flood-influencing factors with CMIP6 remote sensing simulation data.
• Provides accurate predictions of spatiotemporal distribution shifts of future floods under various climate change scenarios (2030, 2050, 2070, 2100).
• Offers critical insights into the most influential factors affecting flood susceptibility, highlighting the significant roles of land use change, elevation, and slope, and challenging traditional assumptions about less impactful factors.
• Establishes a paradigmatic reference for regional flood risk management under global climate change scenarios, demonstrating adaptability across diverse geographical environments and potential for integration with early warning systems.
• Delivers a quantifiable decision-support tool for policymakers and urban planners to implement nature-based solutions (NbS) and develop comprehensive flood governance strategies, considering both natural and socioeconomic dimensions.

Funding

• Not explicitly detailed in the paper.

Citation

@article{Laghari2025Predicting,
  author = {Laghari, Azhar Ali and Shen, Yongheng and Kumar, Akash and Abro, Qurat-ul-ain and Shen, Yanli and Guo, Qingxia},
  title = {Predicting spatiotemporal changes in flood prone regions using PSO-ML coupling under climate change scenarios},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-26939-5},
  url = {https://doi.org/10.1038/s41598-025-26939-5}
}