Hybrid Mixed-Effect Diffusion model (H-MED) for longitudinal air quality analysis


Tanrıverdi İ., Yozgatlıgil C.

ENVIRONMENTAL MODELLING & SOFTWARE, pp.1-24, 2025 (SCI-Expanded)

  • Publication Type: Article / Article
  • Publication Date: 2025
  • Doi Number: 10.1016/j.envsoft.2025.106746
  • Journal Name: ENVIRONMENTAL MODELLING & SOFTWARE
  • Journal Indexes: Applied Science & Technology Source, Scopus, Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), Academic Search Premier, PASCAL, Aqualine, Aquatic Science & Fisheries Abstracts (ASFA), Biotechnology Research Abstracts, Communication Abstracts, Compendex, Computer & Applied Sciences, Environment Index, Geobase, Greenfile, INSPEC, CAB Abstracts, Metadex, Pollution Abstracts, Public Affairs Index, Veterinary Science Database, Civil Engineering Abstracts
  • Page Numbers: pp.1-24
  • Middle East Technical University Affiliated: Yes

Abstract

Air pollution poses severe threats to global public health and environmental sustainability, with complex spatio-temporal dynamics. Traditional air quality models face limitations in capturing longitudinal dependencies and managing computational costs at scale. To address these challenges, we propose the Hybrid Mixed-Effect Diffusion (H-MED) model, a novel framework integrating mixed-effects regression, Gaussian-Process Regression (GPR), Long Short-Term Memory (LSTM), and diffusion-based learning for enhanced predictive accuracy and uncertainty quantification. Applied to a comprehensive longitudinal dataset spanning 61 countries (2013–2023), H-MED demonstrated superior performance, achieving the lowest prediction errors (MAE:0.1423, RMSE:0.1925), while maintaining computational efficiency (5.69s). The model significantly outperformed state-of-the-art approaches including advanced spatiotemporal models such as HITS, DCRNN, ST-GCN, and DeepAR by 15-20\% across all metrics. By incorporating country-level random effects and providing SHAP-based feature attribution, H-MED delivers interpretable, high-resolution forecasts that capture regional pollution heterogeneity and enable targeted policy interventions, a critical advantage over black-box deep learning models that lack explainability for environmental decision-making.