Deep learning based river surface ice quantification using a distant and oblique-viewed public camera

Abstract

Long-term monitoring of surface ice concentration and ice pan properties are of great importance for understanding river freeze-up processes but current data are limited due to challenging winter environments. This study investigated the use of oblique images of river surfaces captured at long focus distances for long-term monitoring of surface ice conditions. Image data from a public camera mounted on a building roof top captured during five freeze-up seasons was used. A deep learning based hybrid image processing algorithm consisting of image classification, rectification, segmentation and extraction of river ice properties was developed to compute surface ice concentrations as well as ice pan size and shape properties. The validity and accuracy of the image data and processing algorithms were evaluated. Time series of multi-year surface ice concentrations as well as size distributions and shape properties of ice pans during freeze-up on the North Saskatchewan River are presented. A lognormal distribution was found to fit the ice pan size distributions for all years. Overall the size and shape of ice pans were relatively stable from year to year. The diameter of ice pans ranged from 0.55 m ∼ 15.03 m with a mean size of 1.82 m. The ice pans are generally elliptical shaped. The daily mean ice pan diameter varied from ∼1 m to ∼3 m. Properties of slushy and crusty ice pans were also estimated by visually separating image periods with dominant ice pan types and it was found that slushy ice pans were generally smaller and slightly less irregular compared to crusty ice pans. This study demonstrates that accurate long-term monitoring of river ice conditions as well as statistical properties of ice pans can be obtained using images captured by a distant and oblique-viewed camera that was not intentionally set up for ice research.