AGL-UNet: Adaptive Global-Local Modulated U-Net for Multitask Sea Ice Mapping

Area of Science:

• Remote Sensing
• Oceanography
• Artificial Intelligence

Background:

• Automated sea ice mapping is crucial for Arctic route planning and climate change studies.
• Increasing satellite data volume necessitates efficient sea ice analysis methods.

Purpose of the Study:

• To develop a multi-task sea ice mapping framework for integrating multi-sensor data.
• To improve the accuracy and efficiency of automated sea ice characterization.

Main Methods:

• Utilized a U-Net architecture with ARC blocks for multi-sensor feature fusion.
• Incorporated a GLCM block for advanced feature modulation.
• Implemented an adaptive loss weighting strategy for multi-task optimization.

Main Results:

• Achieved a 1.33% improvement in the combined score compared to the leading AutoIce Challenge method.
• Increased F1 scores for stage of development (SOD) by 2.85% and floe size (FLOE) by 3.44%.
• Demonstrated effectiveness of proposed blocks and adaptive weighting through ablation studies.

Conclusions:

• The proposed framework effectively handles multi-sensor data for sea ice mapping.
• The method shows significant potential for enhancing ocean environment monitoring capabilities.
• Adaptive multi-task learning balances performance across different sea ice parameters.