Image-Based Outdoor PM2.5 Concentration Estimation Using the BENet-VGG16 Model and Feature Engineering
IEEE Access · 2026 · 14 · 37918-37936
DOI: 10.1109/access.2026.3671821 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Hua-Nan Hsu und 1 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Science and Technology Council (NSTC) of Taiwan, Smart Sustainable New Agriculture Research Center (SMARTer) at NSTC of Taiwan.
Abstract (Originalsprache)
Air pollution remains a persistent environmental challenge, with numerous studies conducted in recent years to address its impacts. Traditional manual sampling methods at monitoring stations encounter limitations such as inconvenient site selection and high operational costs, while sensor-based measurements are often prone to substantial errors. Consequently, recent research has explored innovative approaches that estimate PM2.5 concentrations by utilizing image arrays or extracted features as inputs to neural networks or machine learning models. However, many existing studies fail to automatically exclude defective conditions, lack differentiation between sky and non-sky regions, and rarely incorporate considerations related to data quality control and the interrelationships among regional characteristics. To bridge these research gaps, this study presents and assesses a systematic image-based methodology for estimating outdoor PM2.5 concentrations. The approach integrates feature engineering (e.g., sky image segmentation, weather classification, image preprocessing, feature extraction), the design and configuration of deep learning model architectures, as well as data fusion and estimation strategies. Regarding model architecture, the advantages of employing VGG16-based networks for transfer learning are thoroughly examined, alongside considerations of system resource consumption, training efficiency, overfitting risks, and model parameter scale. The proposed Branch Ensemble Network using the VGG16 (BENet-VGG16) architecture was assessed through 5-fold cross-validation applied to validation datasets collected from monitoring stations located in Renwu, Xitun, Qiaotou, and Linyuan, Taiwan. The optimal average coefficients of determination (R2) achieved were 0.843, 0.836, 0.891, and 0.823, respectively, with corresponding root mean square errors (RMSE) of 6.400, 5.059, 5.839, and $7.310~\mathrm {\mu }$ g/m3. Moreover, the experimental outcomes indicate that the model possesses satisfactory generalization abilities across a range of geographic contexts, while simultaneously achieving parameter reduction, thereby balancing accuracy with computational efficiency.
Förderung
- National Science and Technology Council (NSTC) of Taiwan
- Smart Sustainable New Agriculture Research Center (SMARTer) at NSTC of Taiwan
Zitieren
APA
Hsu, H., & Wen, C. (2026). Image-Based Outdoor PM2.5 Concentration Estimation Using the BENet-VGG16 Model and Feature Engineering. IEEE Access, 14, 37918-37936. https://doi.org/10.1109/access.2026.3671821
Vancouver
Hsu H, Wen C. Image-Based Outdoor PM2.5 Concentration Estimation Using the BENet-VGG16 Model and Feature Engineering. IEEE Access. 2026;14:37918-37936. doi:10.1109/access.2026.3671821
BibTeX
@article{Hsu2026,
author = {Hsu, Hua-Nan and Wen, Chih-Yu},
title = {Image-Based Outdoor PM2.5 Concentration Estimation Using the BENet-VGG16 Model and Feature Engineering},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {37918--37936},
doi = {10.1109/access.2026.3671821}
}
EUVIMED