Theses and Dissertations
ORCID
https://orcid.org/0009-0007-2473-3574
Advisor
Chen, Jingdao
Committee Member
Mittal, Sudip
Committee Member
Chen, Zhiqian
Date of Degree
5-16-2025
Original embargo terms
Visible MSU Only 1 year
Document Type
Graduate Thesis - Campus Access Only
Major
Computer Science (Research Computer Science)
Degree Name
Master of Science (M.S.)
College
James Worth Bagley College of Engineering
Department
Department of Computer Science and Engineering
Abstract
The study of planetary surfaces heavily depends upon space rovers that gather detailed images of terrain needed for analysis and navigation. Deep neural networks and other sophisticated machine learning techniques are necessary for autonomous navigation in challenging terrain. However, the inconsistent annotations by citizen scientists frequently hinder the performance of these models. This study seeks to optimize terrain segmentation to improve the autonomous capabilities of future Mars rovers by presenting a novel weakly supervised learning framework to handle noise and unreliability in datasets. Using factors like number of clicks, pixel accuracy, and annotator dependability, the method utilizes annotation metadata in the training process through a custom weighted cross-entropy loss function. Through extensive data analysis, outliers are excluded and key features are extracted to improve the dataset’s reliability and effectiveness, which lead to more precise training. Thus the segmentation model significantly improves rovers’ autonomous navigation capabilities and advances planetary exploration.
Sponsorship (Optional)
NASA
Recommended Citation
Dutta, Malika, "Optimizing Mars terrain segmentation with weakly supervised learning: A focus on weighted loss from annotation metadata" (2025). Theses and Dissertations. 6480.
https://scholarsjunction.msstate.edu/td/6480
