The snow cover is an important variable because it affects climate, water cycle, and weather. For this reason, GCOS(Global Climate Observing System) determined the snow cover as one of the ECVs(Essential Climate Variables). In addition, GCOS provides requirements for ECVs that scientists are needed to be utilized. The requirements are contained with a spatio-temporal resolution, stability, and uncertainty.
NMSC(National Meteorological Satellite Center) developed an algorithm called SCSI(Snow Cover/Sea Ice) with a 2km spatial resolution to distinguish snow cover and sea ice from other land covers. The SCSI algorithm used the GK-2A(GEO-KOMPSAT-2A) CM(Cloud Mask) product for cloud detection. However, thresholds used in GK-2A CM product are sometimes unstable, causing snow pixels to be classified as clouds.
For this reason, in the SCSI algorithm, the process of determining whether a pixel classified as clouds is snow or not. Therefore, the snow cover detection algorithm that utilizes the characteristics of radiance change by using multi-temporal GK-2A imagery is proposed. Additionally, in order to meet the requirements of GCOS, the pansharpening method was used to generate GK-2A imagery with 500m spatial resolution. Since the fused image should have similar spectral characteristics and spatial details to the original image, the optimal method for the GK-2A image was found through qualitative and quantitative analysis.
In order to ensure stable accuracy of snow cover detection, the performance of the cloud mask algorithm is essential because the spectral characteristic between snow and cloud are similar. Therefore, the adaptive thresholds algorithm for cloud detection was developed considering the advantage of geostationary satellite and the difference in spatial/spectral characteristics between snow and cloud. The misclassified snow pixels were removed with NDVI(Normalized Difference Vegetation Index), near-infrared and thermal infrared bands in the snow cover detection part. After the cloud mask is applied and misclassified snow cover pixel is removed, the NDSI(Normalized Difference Snow Index) images observed during the day are stacked up. The imperceptibly remaining clouds were removed by applying a threshold, and daily snow cover data was finally generated.
For qualitative and quantitative verification, various satellite images and level-2 products, which are GK-2A SCSI, CALIPSO(Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations), PlanetScope, Landsat-8, IMS(Interactive Multisensor Snow and Ice Mapping System), and VIIRS(Visible Infrared Imaging Radiometer Suite) were used. As a result, in comparison with the Landsat-8-based snow cover product, the proposed algorithm showed an F1 score of 0.93, which was 0.13 higher than GK-2A SCSI and 0.03 higher than the VIIRS snow cover product. In addition, the comparison with CALIPSO-based cloud product showed satisfactory accuracy.