Uncertainty-aware Unsupervised Machine Learning to Draw Coastline

Abstract

Automatic drawing of coastlines with satellite imagery is a crucial factor in detecting coastline shifts due to global climatic changes. However, unavailability of labelled information poses a challenge. We propose an explainable unsupervised machine learning model to automatically draw coastlines in the Baltic sea area to create a ‘pre-labelled’ dataset, which clearly delineates the boundary pixels between sea and land. Model uncertainty is computed for each pixel and communicated to the domain experts for verification. The domain expert rectifies any error made by model with an interactive tool for human-ML interaction. Initially, we used only Sentinel-2 imagery which had cloud-related issues, later, we have proposed an uncertainty-based approach to fuse Synthetic Aperture Radar images with Sentinel-2 images.The final results show greater accuracy and less uncertainty. An user-interface tool is also presented to validate the segmentation results and integrate human expert’s knowledge