As brain imaging data such as fMRI is growing explosively, how to reduce its size but not to lose much information becomes a pressing problem. To address this problem, this work aims to represent resting state fMRI (rs-fMRI) signals of a whole brain via a statistical sampling based sparse representation. Specifically, we improve the online dictionary learning and sparse coding algorithm by adding a sampling step before the whole-brain sparse representation. Our comparison experiments demonstrated that this sampling-enabled sparse representation method can speedup by ten times without losing much information. In particular, our results showed that anatomical landmark-guided sampling is substantially better than statistical random sampling in reconstructing concurrent functional brain networks from the Human Connectome Project (HCP) rs-fMRI data.