Loading...

Developing A Machine Learning Framework for Predicting Fine Dust in the Lake Urmia Basin

Khaleghi, Hadiseh | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58561 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Sheikholeslami, Razi
  7. Abstract:
  8. In recent decades, the unprecedented decline of water resources has turned Lake Urmia into one of Iran’s most critical environmental hotspots. The expansion of the lake’s dried bed has intensified the occurrence of salt dust storms, leading to serious consequences for public health, agriculture, the local economy, and ecological sustainability across the region. The main objective of this study was to develop a data-driven model to predict PM₁₀ concentrations over the Lake Urmia watershed using advanced machine learning algorithms. To achieve this, a comprehensive dataset was compiled, including meteorological variables, land cover types, remote sensing indices such as Aerosol Optical Depth (AOD), and a variety of spatial and temporal features. Several models were trained using algorithms such as Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), XGBoost (XGB), LightGBM (LGB), and an ensemble method based on Stacking. The performance of these models was evaluated through both cross-validation (2000–2020) and out-of-sample testing (2021–2023) using statistical indicators including R², MAE, and RMSE. The findings revealed that the Stacking model, which integrates multiple boosting-based algorithms, outperformed all other models across all evaluation metrics. It demonstrated high predictive accuracy and strong capability in capturing both temporal fluctuations and spatial variability in PM₁₀ concentrations. In particular, this model showed exceptional performance in reconstructing the dynamic behavior of pollutants during highly variable periods and spatial hotspots. Structural analysis of the models indicated that variables such as AOD, air temperature, temporal factors (e.g., month and season), spatial location (e.g., latitude), and several hydrological variables, including soil moisture and surface pressure, played significant roles in improving model performance. In contrast, some variables such as topographic features had a lesser impact in the tested models, although their influence may vary depending on the study’s spatial scale, geographic context, and model configuration, and should be further explored in future research. Overall, the results underscore the importance of using hybrid and boosting-based approaches for modeling complex environmental data. They highlight that intelligent model design, careful selection of influential variables, and comprehensive error analysis can provide a solid foundation for the development of effective air quality monitoring and early warning systems at both local and regional scales
  9. Keywords:
  10. Dust ; Lake Urmia Watershed ; Machine Learning ; Ensemble Learning ; Particulate Matter Less than 10 mm

 Digital Object List

 Bookmark

No TOC