Linas Bevainis, Martynas Bielinis, Agimantas Česnulevičius, Artūras Bautrėnas

Lithuanian river ice detection and automated classification using machine-learning methods

Abstract In regions susceptible to river freezing and flooding, river ice detection is a priority. Localization of ice jams and ice drift zones could mean a faster and better response to possible flooding areas, and classification of river ice could help better predict freezing and thawing conditions that hinder the use of commercial and recreational river transport. As many freezing-prone rivers are located in regions with short winter days and common cloud cover, the use of optical sensors can be very limited, therefore, the use of Synthetic Aperture Radar (SAR) – a microwave imaging radar – is more applicable. In this article, Sentinel-1 SAR C-band imagery is used to create derivate texture rasters, which are analyzed, compared with known optical imagery and then considered for river ice detection and discrimination. These results are compared in terms of their effectiveness for river ice discrimination, and the most useful methods are selected. The chosen methods are then compared in an experimental machine-learning model capable of detecting and classifying ice and water. Various machine-learning approaches (both classical and deep-learning) are considered and compared, and the best models are selected. The purpose of this research is to analyze the capability of texture rasters, calculated from a gray-level co-occurrence matrix (GLCM), to discriminate river ice. Texture rasters have recently been applied for river ice classification by de Roda Husman et al. (de Roda Husman et al. 2021), but included only three metrics. This research aims to expand on this knowledge by comparing eight metrics instead of three, as well as including an experiment with a deep-learning model. The results demonstrate that in machine-learning experiments, only one texture measure out of eight (GLCM Mean calculation) is able to discriminate river ice better than discrimination from a standard SAR backscatter intensity image (the baseline).

Doi https://doi.org/10.5200/baltica.2023.1.1

Keywords Sentinel-1; SAR; deep learning; classification methods

Full text