Safe and efficient terrain traversal in natural environments can be improved with robust semantic segmentation models capable of handling complex outdoor scenes. In this study, we evaluate the performance of recent deep learning-based segmentation methods, specifically transformer-based architectures fine-tuned on the ADE20K dataset, within the same outdoor area comprising a mix of urban and natural scenes. The most robust method from this set is subsequently fine-tuned to enhance the distinction between two ground-level classes, which are critical for accurate outdoor scene understanding and for enabling safe terrain traversal in natural settings. Given the high performance achieved by the resulting approach, we adapted the original method to perform semantic segmentation of forest environments, based on a well-defined set of semantic classes designed to support autonomous navigation tasks in such contexts. Two different versions were developed: one trained on the GOOSE dataset, and another trained on both the GOOSE and CWT datasets. These models were then validated using a real-world forest dataset. Our findings highlight the key strengths and limitations of all models evaluated, offering insights into their suitability for outdoor scene understanding, particularly in forested environments. We also discuss the lessons learned and propose strategies to enhance segmentation robustness and efficiency, with a focus on enabling safe terrain traversal in forest settings.