# Bayesian U-Net for Segmenting Glaciers in SAR Imagery

**Authors:** Andreas Hartmann, Amirabbas Davari, Thorsten Seehaus, Matthias Braun, Andreas Maier, Vincent Christlein  
**arXiv:** [2101.03249](https://arxiv.org/abs/2101.03249)  
**Date:** 2021-01-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2101.03249v2/

## Key points

- Study focuses on using Bayesian U-Net for segmenting glaciers in SAR imagery
- Emphasis on monitoring calving front fluctuations impacting ice flow within glacier systems
- Exploration of convolutional neural networks for automated glacier segmentation due to challenging manual delineation of SAR images
- Proposed approach involves computing uncertainty and integrating it into an Uncertainty Optimization regime as a two-stage process
- Creation of a probabilistic Bayesian Neural Network by incorporating dropout as a random sampling layer in U-Net architecture
- Estimation of model uncertainty for each pixel in the segmentation mask through multiple forward passes
- Inclusion of an uncertainty map enhances segmentation accuracy and provides guidance for manual data annotation experts
- Significant improvement in segmentation performance achieved by feeding the uncertainty map back into the network, with a Dice similarity of 95.24%
- Comparative evaluation shows superior performance, especially in image A and B where Uncertainty Optimization yields optimal results
- Bayesian U-Net approach proves effective in accurately segmenting glaciers from SAR imagery and advancing automated glacier monitoring techniques
- Importance of incorporating uncertainty estimation highlighted to enhance segmentation outcomes and improve model robustness and accuracy for complex environmental applications like glacier monitoring

## Summary

The study focuses on the use of Bayesian U-Net for segmenting glaciers in SAR imagery, with a specific emphasis on monitoring the calving front fluctuations that impact ice flow within glacier systems. Manual delineation of SAR images is challenging and subjective, prompting the exploration of convolutional neural networks for automated glacier segmentation. The proposed approach involves computing uncertainty and integrating it into an Uncertainty Optimization regime as a two-stage process. By incorporating dropout as a random sampling layer in a U-Net architecture, a probabilistic Bayesian Neural Network is created to estimate model uncertainty for each pixel in the segmentation mask through multiple forward passes. The inclusion of an uncertainty map provides valuable guidance for experts in manual data annotation, enhancing segmentation accuracy. Notably, feeding the uncertainty map back into the network results in a significant improvement in segmentation performance, achieving a Dice similarity of 95.24%. Comparative evaluation with deterministic U-Net-based pipelines demonstrates superior performance, particularly in image A and B where Uncertainty Optimization yields optimal results. Overall, the Bayesian U-Net approach proves effective in accurately segmenting glaciers from SAR imagery and showcases its potential for advancing automated glacier monitoring techniques. The study highlights the importance of incorporating uncertainty estimation to enhance segmentation outcomes and underscores the value of probabilistic modeling in improving model robustness and accuracy for complex environmental applications like glacier monitoring.