This paper presents a novel algorithm that estimates energy consumption of exploration robots for the efficient mobility in natural terrain. The energy estimation method is beneficial to the energy management problem to establish the robot autonomy, especially in energy-limited environments such as planetary surfaces and isolated volcanoes. The key idea of the proposed approach is to employ the terrain classification into the dynamics-based energy estimation model, so that it can adapt to the variability of terrain properties. A vibration-based terrain classifier is proposed in this paper, which analyzes vibration signals in the time-frequency domain and learns terrain patterns by a supervised learning technique. The terrain classification results are used to determine terrain-dependent parameters in the energy estimation model. A field test has been conducted in a volcanic field to show the validity of the proposed algorithm. The proposed method successfully demonstrates the capability to estimate energy consumption, while it raises discussions about determining terrain types in real natural terrain.