Identification Of Targeted Regions On An Analogue Site Of The Moon By Using Deep Learning Segmentation Algorithm

In this study, a possibility in use and performance of a deep-learning segmentation algorithm are investigated for automatically identifying targeted local regions, which could be hardly separated by a clear feature line. The 4 classes targeted in this study are topographically related objects or regions such as small crater, slope of mound, rock, and icy area in different colors, etc. Even if it is not exactly the same as the ones on the Moon in terms of its surrounding environment, their shape and colors even in different scales were tried to mimic as likely as possible. In order to acquire the information, the shape of the object in the image must be accurately segmented individually. The mask region-based convolutional neural network (Mask R-CNN), an open-source deep learning instance segmentation algorithm, has been adopted in this study for that purpose. A labeled data set is composed from the rover images, which has been taken on the analogue lunar. Then, a number of training are undertaken in a set of conditions and verified in a standard manner. It is shown that object identification and regional segmentation are highly workable with good agreement in comparison of the known and predicted regional information. This kind of deep learning application could be further enhanced for providing a key information to support a lunar surface exploration in future.