Semantic Segmentation of Martian Landforms with Sparse Scribble Annotations Using a Pseudo-Labeling Strategy

Abstract

Deep learning-based semantic segmentation techniques play a crucial role in the rapid mapping of Martian landforms. However, creating an annotated dataset for Martian landforms is a labor-intensive and time-consuming process. To reduce the manual effort required for labeling landform masks, this study adopts scribble annotation and develops a weakly supervised semantic segmentation framework that leverages a pseudo-labeling strategy to address the challenge of limited labeling information inherent in scribble annotations. We applied this framework to a self-constructed Martian landform semantic segmentation dataset. The experimental results demonstrate that our weakly supervised approach achieves good semantic segmentation performance in Martian environments and effectively extracts landform masks. This highlights the potential of our method to facilitate efficient and accurate mapping of Martian geomorphic features with reduced annotation effort.