Mars Multipspectral Image Classification Using Machine Learning Techniques

This paper represents a novel application of machine learning techniques for MARS rock detection using multispectral data. The feature set contains spectral data captured from the NASA MER Pancam instruments. The slope features, PCA features, statistic features and features in different colour space derived from the raw multispectral data are also added to the full feature set in order to enlarge the searching range of optimized features. Fuzzy-rough feature selection (FRFS) is employed to generate good feature sets with lower dimension. Some machine larning classification methods (1NN, 5NN, Bayes, SMO and Dtree) and cluster method (FCM) are utilized to classify the rock from soil using the selected feature. The experimental results show that the FRFS can produce a low-dimentional feature set with improved classifying and clustering results thereby enhancing the efficacy and accuracy of rock detection.