GLFFuse: A Multimodal Feature-Level Fusion Network for Multitask Fine-Grained Recognition of Arctic Sea Ice

Accurate monitoring and recognition of Arctic sea ice are essential for understanding global climate change and the evolution of polar ecosystems. With the rapid advancement of satellite remote sensing technologies, integrating data from multiple remote sensing sources has shown strong potential for improving sea ice recognition. However, existing studies have not sufficiently explored how to jointly capture high frequency spatial details and low frequency global structures from multisource observations, nor have they fully addressed effective information interaction across different data modalities. To overcome these limitations, this study proposes a novel multimodal fusion network, termed Global Local Feature Fusion Network (GLFFuse), designed for fine grained Arctic sea ice recognition. The proposed framework integrates Synthetic Aperture Radar (SAR) imagery, Advanced Microwave Scanning Radiometer 2 (AMSR2) data, ECMWF Reanalysis v5 (ERA5) data, and other auxiliary variables. It combines a long short range attention mechanism with an invertible neural network (INN) to jointly model global contextual patterns and local structural details, thereby enhancing the complementarity among multimodal features. Extensive quantitative and qualitative evaluations on the AI4Arctic dataset demonstrate that the proposed feature level fusion strategy consistently outperforms conventional convolutional neural network based and attention based models across different sea ice recognition tasks. In addition, seasonal analysis results indicate that multimodal data fusion significantly improves prediction accuracy and stability under varying seasonal conditions, effectively reducing systematic biases and predictive uncertainty across a wide range of sea ice concentrations.