Through a machine learning technique based on seasonal cycles of sea-ice concentration from satellite data over the last 4 decades, our research shows that four regions are sufficient to best regionalize the Arctic. These regions are mainly organized into latitudinal bands and evolve in time and space. The descriptor proposed to monitor Arctic sea-ice changes is the probability to belong to each region. The probability to belong to the permanent sea-ice regions has decreased by 3.1 % per decade.