The article provides an overview of the products created from Sentinel-1, -2, and -3 satellite data within the project “Development of an Ice Charting Information System (Jääkaardi koostamise infosüsteemi loomine).” These products are used for ice charting and are shared with the public through a web application.
The formation, development, and movement of sea ice are complex processes, and modeling them in the Baltic Sea remains a challenge. Hydrodynamic models describing the Baltic Sea have sought to represent sea ice as accurately as possible through various parameterizations (Rogers & Zieger, 2004; Pemperton et al., 2017; Tuomi et al., 2019). In addition to numerical models, today the main source of ice information is radar satellite data (Radarsat-2, Sentinel-1, etc.). The availability of Sentinel-1 SAR data has significantly improved sea ice monitoring capabilities in the Baltic Sea region. For example, within the Copernicus Marine Environment Monitoring Service (CMEMS), a product for tracking sea ice dynamics has been developed, based on Sentinel-1 EW and AMSR2 data (Karvonen, 2012). An overview of products created for detecting ice edges, ice types, and ice phenomena from radar data is provided in Zakhvatkina et al., 2019.
Several machine learning methods have been tested for ice and water surface detection: linear regression analysis, support vector machines (SVM), gradient boosting, and neural networks (NN) (Zakhvatkina et al., 2019). The initially used support vector method (Zakhvatkina et al., 2017; Hong & Yang, 2018) has since been replaced by gradient boosting and multilayer neural network methodologies (Boulze et al., 2020). The reliability of the latter two methods is higher, as demonstrated by the growing trend in applying deep neural networks for detecting surface features and objects from radar satellite data (Zakhvatkina et al., 2019).
Typically, very high values are obtained for model evaluation criteria in such tasks, even above 95% (Ochilov et al., 2012; Laanemäe et al., 2017). However, when the task includes distinguishing between different ice classes, the results become noticeably weaker (Kaleschke & Kern, 2002; Bogdanov et al., 2005).
To achieve the project objectives, data from the Sentinel satellite series launched under the ESA Copernicus programme were used. All data were obtained from the national satellite data centre ESTHub.
For the visual display of satellite information, multispectral data from the Sentinel-2/MSI and Sentinel-3/OLCI instruments were used. Two colour composites were created from Sentinel-2/MSI data: a) based on RGB channels B04, B03, B01, and b) a false-colour composite based on RGB channels B12, B8A, B04. From Sentinel-3/OLCI data, an ocean colour RGB composite was generated using the Python library Pytroll with the SatPy module (https://pytroll.github.io/). Several workflows for calculating ice parameters have been built on Sentinel-1 synthetic aperture radar (SAR) data. Images from both the narrow (Interferometric Wide, IW) and wide (Extra Wide, EW) scanning modes are used, with polarization modes VV+VH and HH+HV respectively. Calibration and noise removal functions were applied to the Sentinel-1 SAR data, after which the images were projected to L-EST97 and the pixel resolution was reduced to 50 m for IW and 100 m for EW (Figure 1).
(Machine translated from Estonian)
Member of the Estonian NATO Association. Work experience: Mechanical Engineering, team/project management. Machine Learning and Aeronautics during Master's program.
LINKEDIN
