Deep Learning-Based Seasonal Forecast of Sea Ice Considering Atmospheric Conditions

Abstract

The ongoing decline of sea ice in the Arctic has heightened the need for accurate sea-ice forecasts to support environmental protection and resource development in the region and beyond. While deep learning has shown promise in seasonal sea-ice forecasting, most of the existing models overlook the crucial influence of atmospheric factors, thereby limiting their ability to capture the intricate characteristics of the sea-ice system and improve forecast accuracy. To address this deficiency, we propose an attention convolutional long short-term memory ensemble network named Atsicn, which integrates atmospheric factors to enhance the precision of multi-step seasonal sea-ice concentration forecasts. Our findings reveal that Atsicn outperforms state-of-the-art dynamic and statistical models, and demonstrates remarkable reliability in extreme years. Furthermore, the impact of atmospheric factors on sea-ice forecasts exhibits significant seasonality, with a relatively minimal impact on forecasts from March to June, a growing impact from July to October, and a persistent yet diminishing impact from November to February. This study provides a practical approach for seasonal sea-ice forecasts and contributes a new perspective to the understanding of the intricate interplay between sea ice and atmospheric factors.