The ability to navigate autonomously using natural landmarks in unexplored environments is a significant challenge. While numerous techniques can identify predefined objects, only a few are applicable for real-time navigation in unfamiliar terrains. A key issue is the efficient selection of a minimum set of landmarks for localization. As humans, we often rely on natural landmarks in our environment before resorting to GPS, such as a house of a certain color, or a specific establishment. On the moon, these natural landmarks take the form of craters, rocky and flat areas. In this context, the challenge is to train the robot to differentiate stable natural landmarks. On the Moon, the absence of an atmosphere and the uniformity of the dark terrain pose significant challenges for feature detection. Moreover, rovers are constrained by data storage and computational limitations. Hence, for successful exploration, it’s imperative to explore feature comparison and detection methods that can withstand the unique lunar terrain and environmental features while minimizing computational time. In this paper, features detection techniques such as SIFT, ORB, AKAZE, and BRISK, along with deep learning models such as AlexNet, ResNet18, and ResNet50, are evaluated and compared using images from the Yutu-1 rover. Experimental results show that AlexNet and AKAZE are more robust regarding lunar terrain features. Although AKAZE detects fewer feature points than other techniques like SIFT, the feature points are detected and compared with high accuracy and the lowest computational cost making it suitable for obtaining fast and accurate navigation information. AlexNet has better precision and accuracy than the other models, such as ResNet50, and has the lowest computational cost which benefits the real-time navigation. This analysis results are expected to guide the most appropriate detection method for autonomous navigation using natural landmarks for future lunar rovers.