Recently, a relatively new and emerging neuroimaging tool, functional near-infrared spectroscopy (fNIRS), has captured the attention of scientific disciplines, especially in designing a brain-computer interface. This technique can provide measures of the hemodynamic changes in the brain. Several attempts have been made to characterize the signals in different conditions. However, they mainly utilized conventional statistical approaches to describe the data. To our knowledge, the matching pursuit analysis of the fNIRS signals has not been examined so far. The current study analyzed and merged the matching pursuit-based indices of oxy-, deoxy-, and total-hemoglobin concentrations to classify mental tasks and rest conditions. It took advantage of a simple subject-specific channel selection methodology and a cascade feature selection approach. Maximum accuracy of 86.2% was achieved by examining different classifiers. In conclusion, the results of the proposed framework sketch promising future directions in the field.