TemporalTemp: Encoding Historical Temperatures for Automated Sea Ice Mapping

This paper investigates the impact of incorporating historical ERA5 reanalysis data into deep learning models for Arctic sea ice parameter prediction. While existing approaches primarily rely on contemporaneous observations, we explore whether encoding temporal variations in ERA5 2-meter air temperature (t2m) can improve classification of sea ice concentration (SIC), stage of development (SOD), and floe size (FLOE). Using a multi-task U-Net architecture, we evaluate different temporal encoding strategies on the AI4Arctic dataset, comparing performance against a baseline model without historical context. Our results show that a 1-year encoding with 4 samples of 90 days each achieves the best overall performance (85.62% combined score), improving SOD and FLOE classification by 1.43% and 1.20% respectively. Qualitative analysis reveals that temperature encoding helps distinguish ice types during transitional seasons but can degrade SIC estimation during melt conditions when air temperatures become misleading indicators. These findings demonstrate that carefully designed temporal encodings can enhance sea ice classification while highlighting the need for complementary data sources to address limitations during melt periods.