In the new millennium, sea ice research has become a hot topic in the Earth observation disciplines, due to its importance for the climate and polar ecosystem. At the same time, it strongly affects a vast amount of anthropogenic activities, including shipping and navigation, oil and gas industry, fisheries, tourism and lifestyle of the indigenous population of Arctic. Therefore, sea ice monitoring becomes a key interest in protecting the Arctic and ensuring safe and efficient activities and polar navigation.

Currently, the most commonly used source of information about sea ice is remote sensing data, especially obtained from synthetic aperture radar (SAR), due to its independence of weather conditions. Nevertheless, sea ice can be monitored by various sensors, using different spectral, temporal, or spatial resolutions. Advanced characterization of sea ice can be achieved by combining relevant information from different sensors in order to obtain useful details about various aspects of sea ice properties. However, it is also true that the additional information may be redundant, corrupted, or unnecessary for the given task, hence increasing the computational complexity of the analysis framework. Thus, we propose a new method to select relevant information from multimodal remote sensing datasets.

Compared to existing dimensionality reduction methods, our algorithm has several fundamental advantages: 1 – It is unsupervised, which means that it is application-independent. 2 – It can be applied to several types of data, at different levels of data fusion. 3 – Since it is applied patch-wise, it emphasizes the particularity and diversity of homogeneous regions. 4 – It exploits a new reversible representation of data that enables a powerful separability without loss of physical meaning. 5 – It employs two similarity measures that account for local and global particularities of the dataset and in turn improves the accuracy of sea ice classification.