Abstract
The Arctic Ocean is undergoing significant changes, leading to a surge in research and commercial activity in the region. As a result, the number of manual observations of ice conditions, known as ASSIST observations, has dramatically increased in recent years. These observations provide valuable information on the state of the Arctic ice cover and have been used as a source of ‘ground truth’ for other satellite products. However, their manual nature raises concerns about the associated uncertainty. To address these concerns, the present study evaluates the uncertainty of total ice concentrations, partial ice concentrations, ice types, and ice thicknesses reported in ASSIST observations. First, the study compares ASSIST observations from 2017 to 2021 against sea ice charts from the Norwegian Meteorological Institute, the US National Ice Center, and the Arctic and Antarctic Research Institute. Second, as a test case, the study evaluates simulated ASSIST observations conducted on optical images by an ice expert, three non-experts, and a deep learning model. Results from the comparisons indicate a good agreement between ASSIST observations and the sea ice charts, with ∼80% of comparisons resulting in a difference of less than ±10% in total ice concentration. Analysis of the remaining 20% reveals high underestimation in summer and overestimation in winter compared to sea ice charts. Groupwise comparison of ice types reveals an accuracy of ∼50% and a good match between the ice thickness reported by ASSIST observations and ice thickness associated with ice types reported by the ice charts, with ∼70% of comparisons resulting in a difference of ±10 cm. Results from the inter-comparison of simulated ASSIST observation point to an agreement between experts and novices in simple scenes with a single type of ice and disagreement in the case of complex scenes.