A Method for Predicting Arctic Sea Ice Concentration Based on Multimodal Feature Fusion and Temporal Trend Analysis

The accurate prediction of Arctic sea ice concentration is essential for polar ecological protection and shipping safety. However, existing prediction methods suffer from insufficient feature representation, which limits their ability to capture the complex spatiotemporal distribution of sea ice. Furthermore, they cannot effectively integrate multi-source, heterogeneous sea ice-related data, resulting in limited prediction accuracy. To address these issues, this paper proposes a Multimodal Feature and Trend analysis (MFT) method for sea ice concentration prediction. In the feature extraction stage, MFT combines a Convolutional Neural Network with a Convolutional Block Attention Module to deeply extract global deep semantic features while also employing the Scale-Invariant Feature Transform algorithm to accurately capture local stable features. To improve processing efficiency for high-dimensional remote sensing data, a coarse-resolution dimensionality reduction strategy is developed to select core spatial features, thereby preserving key spatial distribution information while optimizing computational efficiency. For temporal analysis, the Mann–Kendall (MK) non-parametric test and Sen’s slope method are integrated to quantitatively analyze long-term evolution trends in Arctic sea ice concentration. Experimental results show that the proposed MFT model outperforms random forest (RF), LSTM, and traditional MK methods in both prediction accuracy and computational efficiency.