1. Introduction

The Geostationary Environment Monitoring Spectrometer (GEMS) observes air quality over East Asia from an altitude of approximately 36,000 km (Kim et al., 2020). Currently, GEMS provides 21 fundamental atmospheric composition products, making it a crucial tool for monitoring and analyzing regional air quality. High reflectance elements, such as snow, act as interfering factors in the calculation of radiance values during the analysis of atmospheric pollutants (Lamare et al., 2020). Snow-covered regions exhibit high reflectance, which can distort the radiative signals of atmospheric pollutants, thereby reducing the accuracy of pollutant concentration and distribution analysis (Chen et al., 2024). To mitigate such distortions and improve the accuracy of air quality analysis, it is essential to precisely detect snow-covered areas and compensate for or exclude these regions during radiance calculations (Yeom et al., 2009).

However, GEMS has a technical limitation due to its reliance on the ultraviolet and part of the visible spectrum, without access to the infrared spectrum, which is more suitable for snow detection. Currently, GEMS lacks a dedicated snow detection algorithm and instead utilizes the near-real-time ice and snow extent (NISE) data provided by the National Snow and Ice Data Center (NSIDC) as a snow flag (NIER, 2020). NISE data, derived at a global scale, has a low spatial resolution of 25×25 km and is updated once daily. In contrast, GEMS operates at a significantly higher spatial resolution of approximately 3.5 km, resulting in considerable differences in spatiotemporal resolution (Cooper et al., 2018).

During the production of GEMS outputs, high reflectance snow pixels are excluded; however, the use of low-resolution NISE data can lead to the over-exclusion of pixels compared to GEMS’s finer resolution. This overestimation of snow-covered areas degrades the precision of GEMS outputs and can lead to over- or underestimation of atmospheric pollutant concentrations. Hence, there is a pressing need for a dataset with a resolution matching that of GEMS for accurate snow area detection.

Recent advances in deep learning-based image processing techniques offer promising solutions to these challenges (Yu et al., 2024). Deep learning models, through their multilayer artificial neural networks, have demonstrated superior accuracy compared to traditional algorithm-based approaches, as well as robust performance under diverse environmental conditions (Baek et al., 2020). However, the absence of suitable datasets for training deep learning models with GEMS data has limited the development of enhanced snow detection capabilities for GEMS.

To address this issue, this study aims to construct a deep learning-based GeoAI dataset for improving GEMS’s snow detection functionality. GEMS Level 1C (L1C) data were selected as the input dataset, followed by preprocessing steps such as re-gridding, normalization, and data format conversion. Labeling was performed by referencing various satellite-based datasets, including GEMS, the GEO-KOMPSAT-2A (GK2A) Advanced Meteorological Imager (AMI), and MODIS snow cover data. A rigorous labeling protocol was established, and geographic information system (GIS) software was used for annotation, followed by cross-validation and rasterization to produce the final label dataset.

2. Main content

This study is composed of two main stages: 1) acquisition and preprocessing of input data and 2) creation of label data. During the input data acquisition and preprocessing phase, GEMS L1C data was selected, acquired, and preprocessed. For label data creation, we collected GEMS snow flag data along with the snow cover data from the GK2A AMI and MODIS and created label data through rigorous labeling standards.

2.1 Study area

Fig. 1 visually represents the study area defined for this research. The study region covers the geographical area observed by the GEMS satellite, encompassing the entirety of East Asia and including regions of perennial snow, such as the Tibetan Plateau. GEMS performs scans by adjusting the spatial coverage of the Earth’s surface to focus on areas with relatively low solar zenith angles based on specific times and months. This approach, designed to maximize the information captured, selects regions with optimal solar angles, thereby enhancing the accuracy of radiance values. The primary scanning modes of GEMS, defined by the ground station, adjust the starting point and width of each scan, resulting in four main modes: 1) Half East, 2) Half Korea, 3) Full Central (FC), and 4) Full West. These modes allow for efficient observations of East Asia and neighboring areas, with perennial snow regions maintaining high reflectance year-round.

2.2 Acquisition and preprocessing of input data

This study selected the L1C data from the GEMS as input data. GEMS L1C data are satellite observation data provided after precise geometric and radiometric corrections, containing radiance values for each spectral band, which offer high-resolution spectral information necessary for snow detection (Cho et al., 2023). These characteristics provide valuable information for training deep learning-based snow detection models. In this study, all radiance values were selected as input data to construct a fundamental dataset for detecting snow-covered areas using only GEMS data. To capture snow-covered areas with prominent spectral responses, we collected GEMS L1C data from November 2021 to January 2022, covering the winter season. By leveraging radiance values across all hyperspectral bands, we aimed to capture the unique spectral responses of snow-covered areas, constructing a dataset capable of effectively identifying high-reflectance snow regions.

Given that GEMS's scan coverage and size vary by observation time, consistent reconfiguration of diverse data to share the same spatial extent and resolution is required. For this purpose, we defined a latitude and longitude range of 5°S to 45°N and 75°E to 145°E to encompass all scan regions, re-gridding to a 0.05-degree grid based on this standard. All observational data were adjusted to a consistent spatial resolution, ensuring uniformity across data collected at different times. The re-gridded data was subsequently normalized to values between 0 and 1 through Min-Max normalization. Min-Max normalization aims to reduce distortions caused by varying data scales during deep learning model training (Yu et al., 2021). By converting extensive data collected from various regions and times to a consistent scale, model stability during training is enhanced. The equation for Min-Max normalization is shown in Eq. 1:

Eq. 1

Xnormalized =X−XminXmax−Xmin

wher X represents the original pixel value to be normalized, an X_min an X_max denote the minimum and maximum values in the dataset, respectively.

GEMS L1C data is provided in NetCDF (.nc) format by default; however, to enhance efficiency in analysis and application, it was converted to GeoTIFF format. Although NetCDF is suitable for storing large multidimensional data, its utility in typical GIS software is limited, which can restrict individual analysis and visualization processes. In contrast, GeoTIFF format supports raster data with location information and coordinate systems, enabling broad application across various GIS software and data processing environments. This conversion facilitated visualization and application across multiple research environments, resulting in a total of 207 input data files.

2.3 Creation of label data

In this study, high-precision label data were generated to facilitate the training of a GEMS-based snow detection model. Fig. 2 illustrates the workflow for label data generation. The process involved several systematic steps, including the selection of reference data for labeling, the establishment of labeling methodologies and guidelines, the execution of labeling using GIS tools, thorough validation, and the final rasterization of the data.

To ensure high accuracy in the label data, this study integrated various satellite data. GEMS and AMI satellite snow detection data, combined with false-color imagery created from GEMS blue band images and AMI’s NIR 1.6 μm and SWIR bands, as well as MODIS data, were used to enhance the accuracy of snow annotations. The false-color imagery was created by combining the blue band (450-500 nm) from GEMS with the NIR 1.6 μm and SWIR bands from AMI. False-color images emphasize the visual characteristics of specific elements by applying non-visible bands. In this study, the GEMS blue band, which provides critical information for detecting snow and clouds, was averaged for use, and AMI values were converted to radiance values using a conversion table. False-color imagery was used in the label data production because snow has high reflectance in the NIR band and low reflectance in the SWIR band, appearing bright in the false-color images and allowing clear differentiation from surrounding ground elements.

The snow detection flag data from GEMS and AMI also played an important role. GEMS provides snow flag values using NISE data, which offers global coverage at a resolution of 25×25 km, updated daily to reflect snow presence from the previous day. AMI provides snow detection data at a finer spatial resolution of 2×2 km, updated every 10 minutes, allowing detailed temporal snow detection (Jin, 2023). Additionally, AMI includes a Data Quality Flag (DQF) with snow detection data, which offers crucial information for evaluating snow detection quality and was used in this study’s label data production to enhance the reliability of snow labeling (Jin et al., 2022).

Due to limitations in ensuring data consistency and accuracy with only GEMS and AMI data, MODIS daily global snow detection data was also used. MODIS data covers the globe at a 0.05-degree interval and classifies snow based on the Normalized Difference Snow Index (NDSI) values ranging from 0 to 100 (Hall et al., 2002). Based on prior studies, regions with an NDSI value above 40 were classified as snow in this study’s annotations (Wang et al., 2008).

To enhance the accuracy of label data, this study established specific labeling methods and guidelines for snow detection. Label data was produced in vector format matching the size of each input data using QGIS. Snow detection labels were constructed accurately by referencing GEMS and AMI data, false-color imagery, and MODIS NDSI values. False-color imagery was especially instrumental in visually distinguishing snow from non-snow areas, aiding in the creation of vector label data. Detailed labeling guidelines were set to ensure consistency, excluding pixels over water bodies with low snow likelihood using water body information from OpenStreetMap. AMI data, provided every 10 minutes, was prioritized in labeling over the daily GEMS and MODIS snow data. The DQF from AMI was referenced to enhance labeling reliability, with only pixels marked as having good quality in the DQF included in the final snow labels. During labeling, any gaps within snow areas were eliminated using GIS tools’ ring addition feature, clarifying the boundaries of snow regions consistently.

The label data, created under these detailed criteria, was further enhanced through cross-validation. In the cross-validation process, additional reviewers re-evaluated snow presence for each pixel to minimize errors in labeling. This strengthened label data accuracy, and the final vector-format label data was rasterized and converted to GeoTIFF format for use.

3. Results

Fig. 3 illustrates examples of the input and label data created in this study. Fig. 3A displays the L1C radiance values captured on January 15, 2022, at 2:45 UTC, while Fig. 3B shows the labeled data, with snow-covered areas marked in light blue. The regions identified as snow-covered span several East Asian countries, including northern Korea, northeastern China, and parts of Russia’s Far East. Notably, the snow-labeled areas generally exhibit lower values in the L1C imagery, and the radiance value boundaries align closely with the labels.

To verify data quality and assess the suitability of the generated dataset for snow detection, we applied a deep learning model. We used a widely adopted U-Net model for training to perform snow detection. U-Net is a deep neural network structure known for its high performance in image segmentation, making it well-suited for detecting specific patterns and segmenting objects in satellite imagery. The U-Net architecture features an encoding-decoding structure: during encoding, the input image is progressively downsampled to extract feature maps, and in the decoding phase, these feature maps are upsampled back to the original resolution to classify each pixel (Ronneberger et al., 2015). Additionally, skip connections between the encoding and decoding stages help preserve fine details in the images. Instead of using the entire dataset for training, we divided images captured in the FC mode—the mode offering the broadest observational coverage into training and evaluation sets. Using the U-Net model, the results on the test data showed an accuracy of approximately 0.88 and an Intersection over Union (IoU) score of around 0.65. These initial results suggest that the labeled data created in this study can be effectively utilized, demonstrating the model’s ability to accurately detect snow-covered regions.

Fig. 4 compares the snow detection results predicted by the model with the label data and snow data collected from other satellites. Fig. 4A to 4E show the label data, the model’s predicted snow results, and snow data from the GEMS, AMI, and MODIS satellites, respectively, for December 27, 2021, at 2:45 UTC. The predicted results align closely with both the label data and snow data from other satellites, accurately detecting snow-covered areas. This indicates a reasonable level of reliability in the dataset and the U-Net model for GEMS-based snow detection. Additionally, while existing GEMS snow data relies on the NISE data with a 25 km resolution, resulting in relatively coarse boundary definitions, the model’s predictions display much finer snow boundaries. In the GEMS data, the snow class accounted for approximately 7.6% of the total pixels. However, in the prediction map generated using the U-Net model, the snow class represented 4.3% of the pixels. This indicates that the application of deep learning techniques enables the analysis of an additional ~3% of pixels, which were previously unavailable for analysis due to the limitations of low-resolution data. However, some misclassifications were observed, such as certain water bodies mistakenly identified as snow-covered and failures to detect snow accurately over parts of Japan. Enhancements to the deep learning model’s structure and the inclusion of data from various scan regions and times are expected to contribute to developing a deep learning-based snow detection algorithm for GEMS. This advancement has the potential to address the limitations of existing satellite observation systems and support the development of snow detection algorithms for future environmental satellites.

4. Conclusion and discussion

This study developed a GeoAI dataset to improve a deep learning-based snow detection model for the GEMS. High-reflectance elements, such as snow, interfere with radiance value estimation, requiring preprocessing to ensure the accuracy of GEMS’s core products. However, GEMS currently lacks an internal snow detection algorithm and instead relies on NISE data. To overcome these limitations, the application of deep learning techniques is essential, and an appropriate training dataset for these models is required. For this purpose, we utilized GEMS L1C radiance data to construct suitable input data and created precise label data by referencing snow cover data from GEMS, AMI, and MODIS.

The dataset developed in this study serves as a foundational training resource, enabling deep learning models to detect snow cover using GEMS data. Experimental results using the U-Net model demonstrated the effectiveness of this dataset, achieving an accuracy of approximately 0.88 and an IoU of around 0.65. These results confirm that the dataset developed here is effective for training in snow detection.

The dataset created in this study can help overcome the existing limitations of GEMS’s snow detection capabilities. It is expected to enhance the snow detection functionality of GEMS and serve as a critical resource for developing deep learning-based snow detection models for future environmental satellites.

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 a grant from the National Institute of Environmental Research (NIER), funded by the Ministry of Environment (ME) of the Republic of Korea (NIER-2024-03-02-008). This research was financially supported by the Institute of Civil Military Technology Cooperation funded by the Defense Acquisition Program Administration and Ministry of Trade, Industry and Energy of Korean government under grant No. 22-CM-EO-02.

Data Availability Statement

The data that support the findings of google drive at https://drive.google.com/drive/folders/14ES1l4kd7YUtAazj0mU2oLixMb4jPF8S?usp=drive_link.

Fig. 1.

Study area observed by the GEMS satellite, covering East Asia, including perennial snow regions such as the Tibetan Plateau. GEMS, Geostationary Environment Monitoring Spectrometer.

Fig. 2.

Process flow for creating label data. AMI, Advanced Meteorological Imager; GEMS, Geostationary Environment Monitoring Spectrometer.

Fig. 3.

Example of input and label data used in this study. (A) Level 1C radiance values from January 15, 2022, at 2:45 UTC. (B) Corresponding label data with snow-covered areas indicated in light blue.

Fig. 4.

Comparison of model-predicted snow detection with label data and snow data from GEMS, AMI, and MODIS satellites on December 27, 2021, at 2:45 UTC. (A) Label data created for this study; (B) model predictions; (C) GEMS snow data based on NISE data; (D) AMI snow data; (E) MODIS snow data. AMI, Advanced Meteorological Imager; GEMS, Geostationary Environment Monitoring Spectrometer; NISE, Near-real-time Ice and Snow Extent.

References

• Baek WK, Jung HS, Kim D (2020) Oil spill detection of Kerch strait in November 2007 from dual-polarized TerraSAR-X image using artificial and convolutional neural network regression models. J Coast Res 102(SI):137–144Article
• Chen S, Xiao P, Zhang X, et al (2024) Improved snow-covered forest bidirectional reflectance model incorporating canopy-intercepted snow and atmospheric effects. IEEE Trans Geosci Remote Sens 62:1–17Article
• Cho Y, Kim J, Go S, et al (2023) First atmospheric aerosol monitoring results from geostationary environment monitoring spectrometer (GEMS) over Asia. Atmos Meas Tech 2023:1–29Article
• Cooper MJ, Martin RV, Lyapustin AI, McLinden CA (2018) Assessing snow extent data sets over North America to inform and improve trace gas retrievals from solar backscatter. Atmos Meas Tech 11(5):2983–2994ArticlePubMedPMC
• Hall DK, Riggs GA, Salomonson VV, DiGirolamo NE, Bayr KJ (2002) MODIS snow-cover products. Remote Sens Environ 83(1-2):1–194Article
• Jin DH (2023) Snow and sea ice cover detection for GeoKOMPSAT-2A/AMI: an advanced multiple method integration (AMMI) model. Ph.D. Thesis, Pukyong National University, 126p
• Jin D, Lee KS, Choi S, et al (2022) An improvement of snow/cloud discrimination from machine learning using geostationary satellite data. Int J Digit Earth 15(1):2355–2375Article
• Kim J, Jeong U, Ahn MH, et al (2020) New era of air quality monitoring from space: geostationary environment monitoring spectrometer (GEMS). Bull Am Meteorol Soc 101(1):E1–E22
• Lamare M, Dumont M, Picard G, et al (2020) Simulating optical top-of-atmosphere radiance satellite images over snow-covered rugged terrain. Cryosphere 14(11):3995–4020Article
• National Institute of Environmental Research (NIER) (2020) Geostationary environment monitoring spectrometer (GEMS) algorithm theoretical basis document. NIER. https://nesc.nier.go.kr/ko/html/satellite/doc/doc.do Accessed 4 Oct 2024
• Ronneberger O, Fischer P, Brox T (2015) U-net: Convolutional networks for biomedical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, 5-9 Oct 2015. Springer, Cham, pp 234-241
• Wang X, Xie H, Liang T (2008) Evaluation of MODIS snow cover and cloud mask and its application in Northern Xinjiang, China. Remote Sens Environ 112(4):1497–1513Article
• Yeom JM, Han KS, Lee GL (2009) The characteristics of visible reflectance and infra red band over snow cover area. Korean J Remote Sens 25(2):193–203
• Yu JW, Park CW, Lee WJ, Lee YM, Kim YH, Jung HS (2024) Dataset for deep learning-based gems asian dust dataset. GEO DATA 6(3):175–185ArticlePDF
• Yu JW, Yoon YW, Baek WK, Jung HS (2021) Forest vertical structure mapping using two-seasonal optic images and LiDAR DSM acquired from UAV platform through random forest, XGBoost, and support vector machine approaches. Remote Sens 13:4282Article

Appendix

Figure & Data

References

Citations

Citations to this article as recorded by