IceMaskNet: River ice detection and characterization using deep learning algorithms applied to aerial photography

Abstract

Dynamic ice processes can significantly affect various river characteristics such as hydraulics, sediment transport, water quality and morphology. River ice can also impede ship navigation and can induce flood hazard. Study of ice processes is thus crucial for understanding rivers in cold regions. These processes vary according to the four different phases of river ice development: formation, progression, recession and breakup. Monitoring and observation of river ice by remote sensing and close-range photogrammetry have recently attracted significant attention from river ice researchers, and the emergence of remotely piloted aircraft systems with onboard cameras has facilitated safe surveying of river ice. Despite all the developments in this field, fast and accurate data acquisition is still very demanding. One of the challenging tasks in data acquisition from aerial imagery is ice detection and classification. This study presents a novel algorithm called IceMaskNet for automatic river ice detection and characterization from aerial imagery. IceMaskNet utilizes an improved version of the Mask R-CNN, a novel Region-based Convolutional Neural Network with an additional mask. The presented deep learning algorithm is able to detect river ice from aerial imagery and characterize it as belonging to one of six different classes: broken ice, frazil slush, ice cover, open water, border ice, or frazil pan. The developed algorithm is tested using data collected on the Dauphin River, in Manitoba, Canada. Aerial photography from several sections of the river with various slopes and bend scales were used to train IceMaskNet to detect, classify and characterize river ice. The presented algorithm detected and classified river ice with average accuracies of 95% and 91%, respectively.