SEATTNET: UNET Enhanced with Squeeze-Excited Attention Gates for Ice-Calving Front Segmentation

Studying the process of ice calving, where ice breaks off from glaciers, ice shelves, or similar formations, is crucial due to its contribution to sea level rise. Accurately segmenting ice-calving fronts in satellite imagery presents significant challenges due to the low visual contrast and sparse pixel representation of these regions. To address these challenges, we introduce the Squeeze Excited Attention Gate Module (SEATTNET), a novel hybrid attention module integrating squeeze excitation within the attention gate combined with a UNET architecture. This integration enhances model sensitivity, ensuring focus on the most relevant features for calving glacier segmentation. UNET-SEATTNET effectively adjusts channel-wise feature responses, concentrating on pertinent spatial ice-calving regions while suppressing irrelevant background noise, thus improving model performance and robustness. By incorporating Gradient-weighted Class Activation Mapping (GRAD-CAM), we gain critical insights into feature prioritization. Our approach reduces average Mean Distance Error (MDE) by 62% compared to the base UNET model. Comprehensive evaluation using other metrics such as Cosine Similarity (CS), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Peak Signal to Noise Ratio (PSNR), and Dice Coefficient demonstrate UNET-SEATTNET's superior performance, advancing the state-of-the-art in ice-calving front segmentation and contributing to glacial monitoring and climate change understanding.