For deep space landing missions, spacecraft are required to identify their expected landing sites autonomously because of the extremely long time delay caused by the distance between the spacecraft and Earth. This identification process is desirable to finish within several seconds by onboard computers with limited calculation performance. Moreover, autonomous identification based on natural features of landing sites are highly recommended in future missions, although some artificial target markers have been used for navigation and control to the landing site in some previous missions. To make fast but reliable identification of landing sites for the automatic task, this research utilizes a deep learning processing for images taken in different light-conditions and altitudes. First, a semantic segmentation model for rocks in terrain images is developed. For robust identification, some improvements are introduced in the semantic segmentation process. Then, to identify the same place in images taken at different altitudes, a comparison algorithm based on triangular shapes is applied. Thus after training, the semantic segmentation model can detect the same place from several images in a relatively short computational time.