Field measurements of suspended frazil ice. Part I: A support vector machine learning algorithm to identify frazil ice particles

Abstract

Field images of in-situ frazil ice particles captured using a submersible camera system called the FrazilCam have proven difficult to analyse due to the presence of suspended sediment particles. McFarlane et al. (2017) accounted for this by subtracting an appropriately-scaled sediment size distribution from the overall size distribution, resulting in an estimate of the size distribution of frazil ice particles. However, this method over-compensated for the effect of suspended sediment particles and completely eliminated certain portions of the size distribution representing ice particles with diameters on the order of ~0.1 mm. In order to process FrazilCam images with greater accuracy, a machine learning algorithm has been trained to classify each individual particle as ice or sediment during image processing, resulting in more accurate size distributions of the frazil ice particles. The methodology used to train and validate the machine learning algorithm is described, and the data previously presented by McFarlane et al. (2017) are reanalysed. This resulted in a decrease in the mean diameters for each deployment reported by McFarlane et al. (2017); however, the overall trends reported remained the same.