Snow and Sea Ice Cover Detection for Geo-KOMPSAT-2A/AMI: An Advanced Multiple Method Integration (AMMI) Model

Abstract

Snow/Sea ice cover are essential climate variables, which are essential for the study of environmental changes and global climate systems in the high latitude and polar regions. Snow/Sea ice cover has a variety of seasonal distribution areas. Continuous snow/sea ice cover is essential for understanding Earth and atmospheric science. Since snow/sea ice cover is mostly distributed in high latitude, polar regions, and alpine regions, satellite-based remote sensing is essential. Sea ice exists only in the polar regions and at high latitudes, and is more mobile than snow cover. Due to the spatial variability of sea ice, accurate detection and monitoring of sea ice is important. In addition, due to the recent melting of sea ice due to global warming, the Arctic route is attracting attention around the world. Rapid detection and monitoring of sea ice is important for the operation of the Northern Sea Route. Snow/Sea ice cover is the most basic product of the microwave and optical sensor mounted on Earth's observation satellites. Optical sensor-based snow/sea ice cover is detected through unique spectral features appearing in visible and short-wave infrared wavelength range. A representative index for snow/sea ice cover detection is the Normalized Difference Snow Index (NDSI). Using this index, the satellite-based operational snow/sea ice cover is mostly detected through the static threshold method. However, snow/sea ice cover has reflectance variability depending on conditions such as density, and solar/viewing angle of snow/sea ice. Since the static threshold method-based snow/sea ice cover does not take into account reflectance variability, potential errors exist. Since snow/sea ice cover is mainly distributed at high latitude and polar regions, the uncertainty of snow/sea ice cover is high, especially for geostationary satellite among optical sensors. However, snow/sea ice cover detection using geostationary satellite is essential because geostationary satellite has a very high temporal resolution with an observation period of up to 10 minutes. Geo-KOMPSAT-2A (GK-2A)/Advanced Meteorological Imager (AMI) is the second generation geostationary satellite of the Korea Meteorological Administration (KMA), with a spatial resolution of 500m to 2km and temporal resolution of 2-10 minutes. The GK-2A/AMI observation area includes areas used for snow/sea ice cover research. Accordingly, this dissertation detected snow/sea ice cover using the Dynamic Wavelength Warning (DWW) method and the IST0 method, which are dynamic threshold methods considering the reflectance variability of snow/sea ice via GK-2A/AMI. GK-2A/AMI based snow/sea ice cover algorithm using DWW and IST0 methods is still used as an algorithm of GK-2A/AMI snow/sea ice cover detection algorithm. Snow/sea ice cover is one of the key error factors that distinguish it from clouds. For this reason, most operational satellite-based snow/sea ice covers are detected in the clear sky area after excluding the cloud area via cloud mask. The use of cloud mask does not consider the test of reclassifying the pixel as snow/sea ice cover when the cloud mask detects actual snow as a cloud. Therefore, this dissertation reclassified snow/sea ice cover through separate tests for discrimination of snow/sea ice cover for cloud areas among cloud masks. Snow/Sea ice cover was reclassified for cloud areas using the dynamic threshold method. In addition, the snow area was further reclassified by applying Machine Learning (ML) models, which are currently actively used in the remote sensing field, to the snow/cloud discrimination. For ML models, Random Forest (RF) and Deep Neural Network (DNN) were used, and the optimal model for snow/cloud discrimination using geostationary satellite data was selected. In this dissertation, snow/sea ice cover was calculated via dynamic threshold methods and ML models, which we named as the Advanced Multiple Methods Integration (AMMI) model. AMMI-based model snow/sea ice cover performed verification with S-NPP/VIIRS snow cover, sea ice cover, and ground observation ASOS in situ observation data. As a result of validation using the S-NPP/VIIRS snow cover, the snow/cloud discrimination performance of the AMMI-based snow cover improved by 13.3% and 7.4% compared to the DWW-based snow cover and the GK-2A/AMI snow cover, respectively. Using ground observation data, the snow detection performance of AMMI-based snow cover improved by 30%, 23.3%, and 22.7% compared to S-NPP/VIIIRS snow cover, DWW-based snow cover, and GK-2A/AMI snow cover, respectively. As a result of validation with the S-NPP/VIIRS sea ice cover, the AMMI-based sea ice cover was 96.16% Probability Of Detection, 0.71% False Alarm Ratio, and 99.48% Overall Accuracy, indicating higher accuracy than the sea ice cover of other geostationary satellites. AMMI model-based snow/sea ice cover integrated with snow/sea ice showed high performance through dynamic threshold and ML model, which replaced the existing static threshold method. High-quality AMMI model-based snow/sea ice cover is expected to provide high-quality input data not only for climate change research but also for weather forecasting models.