River Ice Detection and Classification using Oblique Shore-based Photography

Abstract

River ice processes significantly impact various aspects of river systems, such as hydraulics, sediment transport, water quality, and morphology. Therefore, understanding these processes is essential for cold-region river studies, ship navigation, and forecasting ice-induced hazards. Remote sensing and close-range photogrammetry have gained attention in recent years, thanks to the growing accessibility of affordable photogrammetry devices and advances in computer vision. Despite progress, acquiring fast, accurate, and long-term data remains challenging. This study presents a novel application of IceMaskNet, a river ice detection, segmentation, and quantification algorithm, specifically designed for oblique shore-based imagery. Built on an enhanced version of the instance segmentation algorithm, Mask R-CNN, IceMaskNet for oblique shore-based imagery was trained using 1795 manually annotated images of the Dauphin River. The algorithm demonstrates high accuracy in detecting and segmenting various river ice categories, achieving 90 % detection accuracy and 86 % segmentation masking accuracy. The developed algorithm was applied over a set of four years of oblique shore-based imagery along the Dauphin River. The algorithm was used in a case study to efficiently generate quantitative estimate of different ice classes in a section of the Dauphin river from long-term shore-based monitoring, significantly contributing to our understanding of river ice processes. The study shows the complex nature of river ice processes in the Dauphin River, and highlights the influence of factors such as air temperature, river flow, flow velocity, and river hydrodynamic characteristics.