Revolutionizing Martian Terrain Mapping: Precision Segmentation Through Deep Learning

Planetary mapping, autonomous rover navigation, geological analysis, and mission planning are all based on the Martian terrain segmentation. However, differentiating various surface characteristics is not easy because of the variations in illumination, imbalance in classes, and similarity of the appearance of different terrain types. To overcome these obstacles, this paper suggests a more accurate segmentation model, using an attention-enriched ResNet-50 framework, combined with two explainability methods, i.e., Grad-CAM and Integrated Gradients, to enhance the performance and explainability. The Mars_terrain_classification dataset, which consists of 6,153 high-resolution images of various terrain classes (craters, dunes, impact ejecta, slope streaks) was used to train and test the model. The experimental findings indicate that the proposed framework had an overall accuracy of 92.61, a mean Intersection over Union (IoU) of 84.7, and a total F1-score of 88.48, which is better than baseline models like U-Net and DeepLabV3+ given the same training conditions. In particular, the proposed model has led to a 6.41 percentage points F1-score improvement compared to U-Net and 2.59 percentage points improvement compared to DeepLabV3+, and an 8.7 and 3.2 percentage points IoU improvement, respectively. Class wise analysis revealed that Bright Dune (95.89% F1-score), Crater (91.28%), and other terrain (94.23%) have high scores, whereas Impact Ejecta (76.92%), and Slope Streak (80.62%) have lower scores, suggesting the still existing challenges in minority and visually ambiguous classes. The explainability results showed that Grad-CAM and Integrated Gradients always show geologically meaningful areas, which adds to the transparency and reliability of predictions. The results indicate that the presented framework can offer a correct and sensible solution to the high-precision mapping of the Martian terrain.