STEPNet: A Spatial and Temporal Encoding Pipeline to Handle Temporal Heterogeneity in Climate Modeling Using AI: A Use Case of Sea Ice Forecasting

Abstract

Sea ice forecasting remains a challenging topic due to the complexity of understanding its driving forces and modeling its dynamics. This article contributes to the expanding literature by developing a data-driven, artificial intelligence (AI)-based solution for forecasting sea ice concentration in the Arctic. Specifically, we introduced STEPNet—a spatial and temporal encoding pipeline capable of handling the temporal heterogeneity of multivariate sea ice drivers, including various climate and environmental factors with varying impacts on sea ice concentration changes. STEPNet employs dedicated encoders designed to effectively mine prominent spatial, temporal, and spatiotemporal relationships within the data. It builds on and extends the architecture of vision and temporal transformer architectures to leverage their power in extracting important hidden relationships over long data ranges. The learning pipeline is designed for flexibility and extendibility, enabling easy integration of different encoders to process diverse data characteristics and meet computational demands. A series of ablation studies and comparative experiments were conducted to validate the effectiveness of our architecture design and the superior performance of the proposed STEPNet model compared to other AI solutions and numerical models.