Automated terrain segmentation is an important component of autonomous rovernavigation, particularly in environments where computational resources and system robustness are limited. This thesis investigates how different architectural andtraining-related design choices influence rare-class terrain segmentation performanceon the AI4Mars dataset under constrained deployment conditions. The study focuseson lightweight semantic segmentation models suitable for deployment-constrainedenvironments, with particular attention to the detection of the rare Big Rock terrainclass.A three-stage controlled experiment was conducted to evaluate the effects of input resolution, decoder family, encoder architecture, loss functions, augmentationstrategies, training duration, and deployment-related constraints. Several lightweightsegmentation architectures were compared using the AI4Mars Curiosity NAVCAMdataset. The final model was additionally evaluated under simulated deploymentrelevant constraints, including limits on parameter count, latency, memory usage,and survivability under simulated radiation-induced bit-flip faults.Input resolution and loss-function choice had the largest impact on rare-class performance. The strongest configuration combined DeepLabv3+ with a MobileNetV4-Conv-Small encoder at 512 × 512 input, trained with focal–Tversky loss and basicaugmentation, reaching a Big Rock IoU of 0.446 on the gold expert test set with3.00 M parameters. After adding safeguards against simulated radiation faults, thefinal model recovered its accuracy and stayed within the parameter and memory limits, with baseline CPU latency well under budget.The findings demonstrate that lightweight segmentation models can achieve competitive performance within strict resource and robustness limits when training configuration and architectural choices are carefully optimized. The study also highlights the importance of balancing segmentation accuracy with deployment feasibility in constrained autonomous navigation environments.