Mitigating Class Imbalance and Enhancing Unlabeled Data Extraction in Semi-Supervised Deep Learning for Martian Terrain Segmentation

Xiao, T. and Wang, Q. and Atkinson, P.M. and Tong, X. (2025) Mitigating Class Imbalance and Enhancing Unlabeled Data Extraction in Semi-Supervised Deep Learning for Martian Terrain Segmentation. IEEE Transactions on Geoscience and Remote Sensing, 63. pp. 1-19. ISSN 0196-2892

Text (Tong et al TGARS 2025)
Tong_et_al_TGARS_2025.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (2MB)

Abstract

Semantic analysis of Martian terrain (MT) is essential for understanding the Martian surface and facilitating autonomous rover navigation. Currently, deep learning dominates MT semantic segmentation, but costly and time-consuming label generation hampers its application. To mitigate this, semi-supervised segmentation (SSS) methods have been developed to leverage a large number of unlabeled images to enhance training with a limited set of labeled images. Nevertheless, challenges persist regarding the quality of pseudo-labels for unlabeled images and class imbalance in labeled images. To alleviate these issues, we proposed a novel end-to-end and efficient SSS network, namely, Mars Balance Enhancer (MBE). Specifically, to increase the quality of pseudo-labels, a novel adaptive local data augmentation (ALDA) module was injected in MBE. This module identifies the most challenging regions adaptively to train within unlabeled images, thereby guiding the training of unlabeled images effectively. Meanwhile, to further mitigate the impact of class imbalance in labeled images, the symmetrical–cyclical focal (SCF) loss was introduced, which emphasizes minority classes dynamically during key training stages, enhancing the model’s sensitivity to these classes. We evaluated MBE on six open-source MT datasets: MER, MSL, S5Mars, MarsScapes, TWMARS, and SynMars. The results demonstrated the advantage of MBE over five advanced SSS methods across all MT datasets. Ablation studies validated the effectiveness of the ALDA module and SCF loss of MBE. Furthermore, the SCF loss was found to be more advantageous than five commonly adopted benchmark losses. Similarly, the ALDA module was more accurate compared to the widely adopted augmentation strategy, Adaptive Label-aided (ALa) CutMix. MBE provides a robust and efficient framework that holds promise for advancing future applications in remote sensing and planetary exploration.