Bayesian Unsupervised Machine Learning Approach to Segment Arctic Sea Ice Using SMOS Data

Christoph Herbert, Adriano Camps, Jan Florian Wellmann, Mercedes Vall-llossera
Global warming effects are amplified in the Arctic, and sea ice has been declining with the annual minimum for September 2020 ending up being the second lowest in the 42-year satellite record (NSIDC, 2020). Sea ice governs heat transfer and influences atmospheric circulation, which is particularly important because low-and mid-latitude's climates are closely related to polar climate (Francis & Vavrus, 2012; Overland & Wang, 2010) . Monitoring of both sea ice concentration (SIC), as the fraction
more » ... of sea-ice cover within an observed cell, and sea ice thickness (SIT), are necessary for a consistent determination of sea ice dynamics. Microwave radiometry is independent of daylight and at lower microwave frequency it is mostly unaffected by atmospheric conditions. The emissivity in the microwave spectrum depends on the dielectric properties of sea ice, which are a function of its physical composition including salinity, density, surface temperature, and surface roughness. The collected signal is emitted from a radiating layer which depends on the penetration depth at sensor frequency. A snow layer on top of sea ice influences the radiative properties of sea ice and the energy received by the sensor. The contribution of snow to the emitted signal is less in the lower microwave spectrum, and the separability of surface properties, such as open water and sea ice including SIT, is-in theory-feasible. Abstract Microwave radiometry at L-band is sensitive to sea ice thickness (SIT) up to ∼ 60 cm. Current methods to infer SIT depend on ice-physical properties and data provided by the ESA's Soil Moisture and Ocean Salinity (SMOS) mission. However, retrieval accuracy is limited due to seasonally and regionally variable surface conditions during the formation and melting of sea ice. In this work, Arctic sea ice is segmented using a Bayesian unsupervised learning algorithm aiming to recognize spatial patterns by harnessing multi-incidence angle brightness temperature observations. The approach considers both statistical characteristics and spatial correlations of the observations. The temporal stability and separability of classes are analyzed to distinguish ambiguous from well-determined regions. Model uncertainty is quantified from class membership probabilities using information entropy. The presented approach opens up a new scope to improve current SIT retrieval algorithms, and can be particularly beneficial to investigate merged satellite products. Plain Language Summary Remote sensing techniques are commonly used to provide maps of sea ice thickness (SIT). Methods to obtain these maps are based on the sea ice composition and on the signal measured by satellite. Sea ice Composition is spatially complex and changes during its formation and melting. Currently used data from observations of ESA's Soil Moisture and Ocean Salinity (SMOS) mission depend on several sea ice parameters, which hinders good estimation of almost any specific sea ice parameter. In this work, a new method to combine the information contained in SMOS brightness temperature data is investigated, with the aim to divide the Arctic region into a number of smaller areasso called classes. Useful information about sea ice is contained in the spatial and statistical distribution of SMOS data, which are collected at different incidence angles. The relationship between the observations and the statistical properties of the obtained classes allow an assessment of its degree of separability and uncertainty. How classes change in time is used to estimate their temporal stability. The presented approach can be used to investigate the link between a variety of spatial datasets to improve current SIT products, and it can be applied in other scientific fields. HERBERT ET AL.
doi:10.18154/rwth-2021-03755 fatcat:qs7ph4im6janxlqiugc5pjsupa