Optimization of the k-nearest-neighbors model for summer Arctic Sea ice prediction

• Published in: Frontiers in Marine Science
• Main Authors: Lin, Yongcheng, Yang, Qinghua, Li, Xuewei, Yang, Chao-Yuan, Wang, Yiguo, Wang, Jiuke, Liu, Jingwen, Chen, Sizhe, Liu, Jiping
• Format: Article in Journal/Newspaper
• Language: unknown
• Published: Frontiers Media SA 2023
• Subjects: Ocean Engineering; Water Science and Technology; Aquatic Science; Global and Planetary Change; Oceanography; Arctic; Sea ice
• Online Access: http://dx.doi.org/10.3389/fmars.2023.1260047; https://www.frontiersin.org/articles/10.3389/fmars.2023.1260047/full

The Arctic sea ice plays a significant role in climate-related processes and has a considerable effect on humans, however accurately predicting the Arctic sea ice concentration is still challenging. Recently, with the rise and development of artificial intelligence, big data technology, machine learning has been widely used in the field of sea ice prediction. In this study, we utilized a sea ice concentration dataset obtained from satellite remote sensing and applied the k-nearest-neighbors (Ice-kNN) machine learning model to forecast the summer Arctic sea ice concentration and extent on 122 days prediction. Based on the physical characteristics of summer sea ice, different algorithms are employed to optimize the prediction model. A drift-ice correction algorithm is designed to address the unrealistic drift ice around the sea ice edge, and a distance function combined with the spatial pattern is proposed to enhance similarity detection. Deseasonalized and detrended sea ice datasets and an expanded training library are also utilized to improve model performance. Furthermore, sensitivity analysis reveals a positive impact of net surface heat flux on sea ice prediction. The modified Ice-kNN model outperforms climatological and anomaly persistence predictions, demonstrating its applicability to predicting summer Arctic sea ice. The September sea ice extent hindcasts of the modified Ice-kNN model are compared to a variety of models submitted to the Sea Ice Prediction Network, underscoring its potential to improve predictive skill for Arctic sea ice.