Shin et al. (2025) Machine learning-based retrieval of aerosol size and hygroscopicity using horizontal scanning LiDAR and PM data

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2025
• Date: 2025-12-02
• Authors: Juseon Shin, Juhyeon Sim, Matthias Tesche, Jihyun Yoon, Dukhyeon Kim, Youngmin Noh
• DOI: 10.1038/s41612-025-01276-6

Research Groups

• Division of Earth Environmental System Science, Pukyong National University, Busan, Republic of Korea
• Leipzig Institute for Meteorology (LIM), Leipzig University, Leipzig, Germany
• Hanbat National University, Daejeon, Republic of Korea

Short Summary

This study develops a machine learning-based approach to retrieve aerosol size and hygroscopicity by integrating horizontal scanning LiDAR and in-situ PM data, revealing that coarse hygroscopic aerosols dominate the coastal urban region and significantly impact optical properties despite low mass concentrations.

Objective

• To quantitatively characterize the hygroscopic properties of aerosols by integrating long-term mass concentration data from the Korean National Air Pollution Monitoring Network (AirKorea) with lidar-based remote sensing measurements of aerosol optical properties.
• To introduce a machine learning technique to retrieve the dry volume size distribution (VSD) from Optical Particle Counter (OPC) measurements and use it to calculate dry extinction coefficients, which are then compared with lidar-derived ambient extinction coefficients to infer hygroscopic growth.
• To investigate discrepancies between optical scattering loads from lidar measurements and mass concentrations from in-situ measurements, offering a new methodology for real-time, dynamic classification of aerosol types.

Study Configuration

• Spatial Scale: Coastal urban region of Busan, Republic of Korea. Measurements were performed in the port area of Busan, with instruments co-located within approximately 2.7–5.5 km. The lidar scanning area covers an international trade port, ship terminal, and residential zones.
• Temporal Scale: Long-term measurements from October 2021 to September 2024, with data synchronized to 1-hour resolution.

Methodology and Data

• Models used:
  • Machine Learning: Extreme Gradient Boosting (XGBoost) for dry VSD retrieval, Random Forest for aerosol type classification. Neural Networks (NN), Support Vector Regression (SVR), Decision Tree, Support Vector Machine (SVM), and k-Nearest Neighbors (kNN) were used for comparison.
  • Physical Models: Mie theory for calculating dry extinction coefficients, Multi-section Klett inversion method for lidar-derived ambient extinction coefficients.
  • Statistical: Bi-modal Gaussian distribution for VSD, Gradient descent algorithm for VSD parameter retrieval.
• Data sources:
  • In-situ PM data: Korean National Air Pollution Monitoring Network (AirKorea) using Beta Attenuation Monitor (BAM) for PM10 and PM2.5 mass concentrations (dry conditions).
  • Optical Particle Counter (OPC): Grimm model 1.109, providing particle number size distributions (0.17 to 35.75 µm diameter) and mass concentrations (PM10, PM2.5).
  • Lidar data: Horizontally scanning Smart Lidar MK I (Samwoo TCS Co., Ltd, Pukyong National University) operating at 532 nm wavelength, providing ambient (wet) aerosol extinction coefficients.
  • Meteorological data: Automated Surface Observing System (ASOS) station (KMA) for temperature, relative humidity (RH), and visibility.
  • AERONET data: Pusan National University site for long-term averaged complex refractive index (1.466 + 0.0095i at 532 nm).
  • KMA Dust Storm Warnings: Used for classifying dust events.

Main Results

• The XGBoost model achieved high accuracy (R²=0.98, MAE=0.054) in retrieving dry volume size distribution (VSD) from PM data.
• Aerosol types were classified using Random Forest with 83.4% accuracy, revealing the dominance of coarse hygroscopic aerosols (34% coarse hydrophilic pollution, 27% coarse hydrophobic pollution) in the coastal urban region.
• The hygroscopic growth factor (f(RH)) showed higher values and greater variability at higher RH, particularly for coarse-mode aerosols, attributed to sea-salt aerosols in the coastal environment.
• The wet extinction coefficient (αwet) distinctly increased with rising RH, while dry PM concentration and dry extinction coefficient (αdry) remained relatively stable.
• Clean conditions occasionally showed a sharp increase in αwet despite low PM concentrations, highlighting that a small number of hygroscopic aerosols can significantly impact light scattering under high RH.
• Discrepancies between mass concentration and extinction coefficient measurements were attributed to differences in particle size distribution and hygroscopicity, underscoring the limitations of mass-only assessments for optical properties.

Contributions

• Developed a novel machine learning-based methodology to integrate long-term in-situ PM data with lidar remote sensing for dynamic, real-time characterization of aerosol size and hygroscopicity.
• Provided a quantitative assessment of aerosol hygroscopic properties and their impact on optical behavior over a long period, bridging the interpretational gap between dry mass-based and wet optical measurements.
• Introduced a robust machine learning inversion technique (XGBoost) for retrieving dry volume size distribution from limited mass concentration data, extending VSD estimation to periods without OPC observations.
• Established a supervised machine learning classification model (Random Forest) for aerosol types based on size information, hygroscopicity, and meteorological conditions, which effectively explains variations in aerosol physical and optical properties.
• Demonstrated the critical importance of considering aerosol size and hygroscopic properties, beyond just PM mass concentration, for accurate air quality assessment, visibility prediction, and climate modeling, especially in coastal urban environments.

Funding

• National Research Foundation of Korea (NRF) grant funded by the Korea government (MOE) (No. RS-2025-25411926).
• Ministry-Cooperation R&D Program of Disaster-Safety funded by Ministry of Interior and Safety (MOIS, Korea) (Grant 2023-MOIS-20024324).

Citation

@article{Shin2025Machine,
  author = {Shin, Juseon and Sim, Juhyeon and Tesche, Matthias and Yoon, Jihyun and Kim, Dukhyeon and Noh, Youngmin},
  title = {Machine learning-based retrieval of aerosol size and hygroscopicity using horizontal scanning LiDAR and PM data},
  journal = {npj Climate and Atmospheric Science},
  year = {2025},
  doi = {10.1038/s41612-025-01276-6},
  url = {https://doi.org/10.1038/s41612-025-01276-6}
}