Semantic Segmentation of Glacier Calving Fronts Through Deep Learning and SAR Imagery

Glacier calving fronts are play a vital role in understanding glacier behavior and their impact on sea-level rise. In this work, proposed a novel approach for mapping calving fronts in glaciers using deep learning techniques. Specifically, employ a multi-class U-Net architecture and utilize synthetic aperture radar (SAR) imagery for semantic segmentation of glacier calving fronts. dataset is made up of SAR imagery of seven tidewater glaciers situated in Alaska, Greenland and the Antarctic Peninsula. Preprocess the SAR imagery and their respective labels, encoding them for multi-class segmentation. A multi class U-Net model that has been trained on smaller patches of SAR imagery and their respective segmented regions. Use categorical cross entropy loss and Adam optimizer to train the model. Implement class weighting to deal with class imbalance in an effective manner. The trained model is finally tested against a different test set, and accuracy and mean Intersection over Union (IoU) metrics are calculated. The outcome displays the effectiveness of the propose method to well delineate glacier calving fronts from SAR images. In addition, visualize the inferred calving fronts on the SAR images, giving insights into the spatial distribution of calving fronts over various glaciers. Architecture, Adam optimizer, Deep learning.