1. Introduction

Air pollution has emerged as a pressing global issue, with pollutants accelerating global warming and posing serious threats to human health (Brauer et al., 2016). In South Korea, elevated levels of air pollutants, such as particulate matter and ozone, arise both from domestic sources and from pollutants transported via westerly winds from China, leading to significant environmental and health impacts. Monitoring air quality in neighboring regions, particularly in China, has thus become crucial for anticipating changes in South Korea’s air quality and for formulating effective response strategies (Choi et al., 2019; Jang and Yeo, 2015). Industrialized urban areas in China, which emit substantial quantities of pollutants due to industrial activities and vehicle traffic, demand consistent and efficient monitoring. However, China’s vast area of 9,596,960 km²-approximately 95 times larger than South Korea’s 100,210 km²-poses notable challenges in detecting and monitoring sources of pollution.

Satellite imagery-based remote sensing techniques offer valuable tools for acquiring information about inaccessible regions and large-scale phenomena. Due to their broad observational range, satellite data are particularly suitable for detecting extensive pollution sources, such as those in urbanized regions. In recent years, artificial intelligence (AI) techniques have increasingly been applied to satellite imagery analysis, with methods such as object detection, change detection, classification, and semantic segmentation advancing through deep learning approaches that frequently surpass traditional algorithms in performance (LeCun et al., 2015).

This study developed an AI dataset for urban area segmentation to detect sources of air pollution, utilizing 30-m resolution Landsat 8/9 satellite imagery from 2023 and 2024, as well as data from the Geostationary Environment Monitoring Spectrometer (GEMS) and air quality monitoring network. In addition to optical imagery, GEMS and air quality monitoring data, which are used for detecting areas with high concentrations of air pollutants, were incorporated into the input data to construct a more effective dataset for air pollution source detection. A U-Net model, recognized for its high efficacy in semantic segmentation, was then applied to analyze the dataset, with an assessment conducted on its potential utility as a foundational tool for detecting and monitoring urban air pollution sources.

2. Methodology

2.1 Study data

To construct an AI dataset for large-scale urban area segmentation, this study employed imagery from Landsat 8/9, the GEMS, and air quality monitoring network data. The Landsat program, operated by the United States Geological Survey, provides essential information on terrestrial characteristics, including land cover, vegetation, and water resources. Landsat imagery has been widely applied in diverse fields such as land use and land cover change detection, natural resource management, forestry, and disaster response. Launched in February 2013 and September 2021, Landsat 8 and Landsat 9 continue to supply high-resolution multispectral imagery across 11 spectral bands with spatial resolutions of 15 m, 30 m, and 100 m. In this study, Level 2 Landsat data were utilized, which include radiometric, geometric, and atmospheric corrections, with the blue, green, red, and near infrared bands (bands 2, 3, 4, and 5) serving as input data (Table 1).

The GEMS, launched in 2020, is the first geostationary satellite dedicated to the continuous observation of climate-forcing and air pollutant substances across East Asia, including the Korean Peninsula (Choi et al., 2018). GEMS provides critical data on atmospheric constituents essential for environmental monitoring, including aerosols, ozone (O₃), surface reflectance, cloud properties, and nitrogen dioxide (NO₂). This study utilized the Level 4 GEMS dataset, specifically the monthly average estimates of ground-level NO₂ concentrations.

In addition to Landsat and GEMS satellite imagery, air quality monitoring network data were incorporated to refine urban area segmentation. Notably, the concentrations of sulfur dioxide (SO₂), carbon monoxide (CO), and NO₂ provided by the air quality monitoring network are closely linked to urban air pollution sources, making them valuable for identifying key pollutant emission hotspots (Wei et al., 2023). For South Korea, air quality monitoring network data were sourced from AirKorea, which enables data selection based on time period, region, monitoring network, and station locations. Monthly mean concentrations of SO₂, CO, and NO₂ were employed for this study. For Chinese data, station locations were verified using the UNEP Air Quality Monitoring Platform, and corresponding pollutant measurements were obtained from the Air Pollution in World database.

Fig. 1 illustrates the Coverage area of Landsat 8/9 imagery utilized in this study. The imagery collection area for constructing the AI dataset for urban area segmentation includes regions in South Korea and key urban and industrial zones in China, such as Beijing, Tianjin, Hebei, Shandong, Shanghai, Zhejiang, and Jiangsu. These regions in China were selected due to their severe air pollution problems, which stem from high population density, extensive industrial complexes, and increased traffic volumes (Jeon and Kim, 2015). A total of 55 Landsat 8/9 images were acquired between March 2023 and June 2024, comprising 10 images from South Korea and 45 from China. Details of the acquisition regions, dates, and corresponding satellite imagery are provided in Table 2. All imagery was manually inspected to obtain raw data without applying additional cloudcover filtering options during the search period.

2.2 AI dataset construction

The AI dataset for semantic segmentation of satellite imagery consists of paired input and label data. In this study, the input data include satellite imagery from Landsat 8/9, GEMS, and air quality monitoring network data, while the label data comprise ground truth classifications specifically developed for urban area segmentation. Each satellite image in the input data is paired with a corresponding label that provides the correct segmentation class for the urban region.

Label data construction for urban area segmentation was conducted using the open-source QGIS software, following a rigorous data annotation process. A clear and precise set of criteria was established to delineate urban areas, documented in a comprehensive data construction guideline. Urban areas were defined as regions where buildings serve residential, commercial, or industrial functions, encompassing structures such as residential buildings, commercial and industrial facilities, and transportation infrastructure. Non-urban regions, including forest areas, water bodies, and agricultural land, were excluded from the urban class; trees and grasslands within urbanized zones were also excluded.

To enhance objectivity and consistency in label generation, reference data were incorporated to guide urban area delineation, acknowledging limitations inherent in using only Landsat 8/9 imagery. For South Korea, mid-level land cover maps served as reference data, while for China, open-source geographic data from OpenStreetMap and ESA WorldCover 2021 were utilized. South Korea’s Ministry of Environment provides a scientifically classified land cover map that categorizes surface features into 22 classes, supporting a systematic approach to delineating urban areas.

Following these guidelines, urban segmentation labels were initially constructed in vector format. These vector labels were then converted into raster format (TIFF) through a rasterization process to ensure compatibility with the input data. The Landsat 8/9 satellite data and rasterized labels were partitioned into 512×512 pixel patches with a 25% overlap, enhancing spatial detail for model training. Additionally, GEMS and air quality monitoring network data were segmented into 64×64 pixel patches, enabling consistent alignment across different data sources. Fig. 2 illustrates the AI dataset construction process of this study.

3. Results

Through the dataset construction methods described above, a total of 5,000 samples were generated for training. Table 3 summarizes the patch size, data format, and quantity for each component of the dataset. The input data include imagery from Landsat 8/9, GEMS, and air quality monitoring network data, while the label data serve as ground truth for semantic segmentation. Fig. 3 presents representative examples from the constructed AI dataset: Fig. 3A-C show sample inputs from Landsat 8/9 imagery, GEMS imagery, and air quality monitoring network data, respectively, while Fig. 3D illustrates the corresponding label data used as ground truth for these inputs.

To validate the efficacy of the constructed AI dataset for urban area segmentation, a U-Net-based model was employed due to its proven effectiveness in semantic segmentation tasks (Du et al., 2020; Long et al., 2015). The conventional U-Net architecture, which employs a single encoder-decoder structure, was modified to adopt a multi-encoder design, enabling the integration of the three distinct input data types-Landsat 8/9, GEMS, and air quality monitoring network-within a unified framework.

The dataset was divided into training, validation, and test subsets with an 80%, 10%, and 10% split, respectively. The training dataset was used to optimize the model, while the validation dataset facilitated hyperparameter tuning and overfitting assessment. The test dataset, held independent from the training and validation stages, was utilized to objectively evaluate the model’s final performance. Additionally, data augmentation techniques were applied to the training dataset to enhance the model’s generalization ability (Baek et al., 2022).

The model achieved a mean Intersection over Union (mIoU) of 0.8592 and a pixel accuracy of 0.9433. For urban area classification specifically, the model demonstrated a precision of 0.9107, recall of 0.9383, and F1 score of 0.9243, reflecting strong segmentation performance. Fig. 4 illustrates example outputs from the trained modified U-Net model, showcasing predictions for different test samples. Each row in Fig. 4 corresponds to a distinct test image, with columns representing the Landsat 8/9 input imagery, label data, and the model’s predicted segmentation result.

A qualitative assessment of the results revealed that while occasional false negatives and false positives were present, the model generally achieved accurate segmentation for urban areas. For cases with well-defined boundaries, such as those depicted in Fig. 4A, B, the model’s predictions closely aligned with the label data, demonstrating high efficacy in distinguishing between urban and non-urban areas. Conversely, Fig. 4C represents a more complex scenario with ambiguous boundaries between urban and non-urban regions, resulting in minor discrepancies between the predicted segmentation and the label data.

4. Conclusion

This study developed a GEO AI dataset tailored for urban area segmentation to facilitate the detection of air pollution sources, integrating Landsat 8/9 satellite imagery, GEMS satellite imagery, and air quality monitoring network data. The dataset construction process was meticulously documented, ensuring that the resulting dataset is well-suited for semantic segmentation tasks in urban area segmentation. Notably, the constructed AI dataset incorporates not only optical imagery but also diverse data types, underscoring its significance in urban area segmentation for identifying air pollution sources. Furthermore, the dataset demonstrates substantial potential as a foundational resource for detecting and monitoring major sources of air pollutants.

In experiments utilizing a modified U-Net model, the dataset yielded robust results, achieving a mIoU of 0.8592 and pixel accuracy of 0.9433, thus verifying its practical applicability. The dataset can be further refined by incorporating data from additional time periods and by including diverse types of air pollutants, thereby enabling more detailed analysis of pollution sources and affected regions. The spatial information of air pollution sources provided by the dataset plays a pivotal role in understanding the movement and distribution of air pollutants. As a result, this GEO AI dataset is expected to play a crucial role in air quality monitoring and pollution source detection systems across South Korea and East Asia. In the long term, it holds promise as a key resource for policy development aimed at air pollution mitigation and environmental management, contributing to sustained improvements in air quality.

Notes

Conflict of Interest

On behalf of all authors, the corresponding author states that there is no conflict of interest.

Funding Information

This work was supported by the 2023 sabbatical year research grant of the University of Seoul.

Data Availability Statement

The dataset supporting the findings of this study is currently under embargo. It is scheduled for public release on AI Hub in April 2025, at which time it will be assigned a DOI and made fully accessible. This approach ensures compliance with data-sharing requirements and facilitates reproducibility and further research.

Fig. 1.

Study area. Coverage area of Landsat 8/9 imagery in South Korea and China.

Fig. 2.

Workflow of artificial intelligence (AI) dataset construction.

Fig. 3.

Examples of constructed AI dataset for urbanized area segmentation. (A) Landsat 8/9. (B) GEMS. (C) Air quality monitoring network data. (D) Label. AI, artificial intelligence; GEMS, Geostationary Environment Monitoring Spectrometer.

Fig. 4.

Examples of constructed artificial intelligence dataset for urbanized area segmentation. (A, B) The model demonstrates accurate segmentation in urban areas with well-defined boundaries, closely matching the label data. (C) In areas with ambiguous boundaries, minor discrepancies are observed between predictions and label data.

References

• Baek WK, Lee MJ, Jung HS (2022) The performance improvement of U-net model for landcover semantic segmentation through data augmentation. Korean J Remote Sens 38(6):1663–1676
• Brauer M, Freedman G, Frostad J, et al (2016) Ambient air pollution exposure estimation for the global burden of disease 2013. Environ Sci Technol 50(1):79–88PubMed
• Choi J, Park RJ, Lee HM, et al (2019) Impacts of local vs. transboundary emissions from different sectors on PM2.5 exposure in South Korea during the KORUS-AQ campaign. Atmos Environ 203:196–205
• Choi WJ, Moon KJ, Yoon J, et al (2018) Introducing the geostationary environment monitoring spectrometer. J Appl Remote Sens 12(4):044005Article
• Du G, Cao X, Liang J, Chen X, Zhan Y (2020) Medical image segmentation based on U-Net: a review. J Imaging Sci Technol 64(2):020508Article
• Jang KS, Yeo JH (2015) The effects of Korean and Chinese economic growth on particulate matter in Korea: time series cointegration analysis. J Environ Policy Adm 23(1):97–117Article
• Jeon SH, Kim YP (2015) A study on the smog reduction strategies in China. Part Aerosol Res 11(3):63–75
• LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436–444ArticlePubMedPDF
• Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, 7-12 Jun 2015, pp 3431-3440
• Wei J, Li Z, Wang J, Li C, Gupta P, Cribb M (2023) Ground-level gaseous pollutants (, SO₂, and CO) in China: Daily seamless mapping and spatiotemporal variations. Atmos Chem Phys 23(2):1511–1532Article

Appendix

Figure & Data

References

Citations

Citations to this article as recorded by