ABSTRACT
Rivers in cold regions such as Canada are covered by ice during a significant portion of the year. River ice processes such as ice cover formation, progression, recession, and break-up can affect river hydraulics, sediment transport, and morphology. Freeze-up is one of the most significant stages of ice cover formation in cold seasons. Given the lack of a reliable predictive model, in situ field investigations, although difficult and potentially dangerous, are crucial for understanding ice cover formation dynamics. Remote sensing and the application of digital cameras in river ice processes monitoring have recently been used by several researchers. The acquired images and video files have been used in several studies for qualitative assessment; however, accurate quantified data acquisition is still very demanding. The emergence of the use of remotely piloted aircraft system in earth siences has facilitated safe river ice observations. Automatic detection of river ice is one of the challenging tasks in this procedure. The main goal of this study was to present a novel approach for river ice detection in an automated pipeline for aerial monitoring of river ice. Aerial imagery were captured during a series of drone flight campaigns along the Dauphin River, Manitoba, Canada. Some of these images were used in this study to train Convolutional Neural Network (ConvNN) models to detect and classify river ice. The trained ConvNN models were then used to detect and classify river ice in the remainder of the aerial images. Results suggest that ConvNN can be used for a better estimation of river ice concentration along the monitored sections of the river.
