Recognition of the Martian minerals based on the deep belief networks method: Application in the CRISM images

In order to decrease recognition inaccuracies of the different minerals with similar single absorption peak by means of the spectral characteristic parameter methods which are difficult to estimate the spectrum of the whole wavelength range, this paper applied the deep belief networks (DBN) method to detect the Martian minerals from the hyperspectral images of the compact reconnaissance imaging spectrometer for Mars (CRISM). According to the method, firstly, the unsupervised layer-by-layer greedy algorithm is adopted to train each restricted Boltzmann machine (RBM) for the sake of learning parameters and extracting the spectral features of the minerals with a single bottom-up pass. Then, it takes advantage of the back propagation (BP) algorithm to tune the parameters learned in the train step and automatically identify the Martian minerals with coupling a suitable classifier. In this paper, the ratios of the minerals spectral and the dust spectral are utilized to identify the mineral samples for sake of decreasing the dust effect. Finally the influences of the sample size, the number of the hidden layer nodes, and the network depth are investigated to established the optimal deep belief networks for the recognition of the Martian minerals. As illustrated by the case of the Mg/Fe smectites and the chlorides from the CRISM images, the experimental results indicate that the recognition accuracy of the DBN method is more than 85%. In conclusion, the DBN method has a better performance in detecting some pixels of the minerals that the spectral parameter algorithm cannot detect in the CRISM images, and the deep learning method could be utilized in the recognition of the Martian minerals automatically.