Arterial spin labeling (ASL) perfusion MRI provides a non-invasive way to quantify regional cerebral blood flow (CBF) and has been increasingly used to characterize brain state changes due to disease or functional alterations. Its use in dynamic brain activity study, however, is still hampered by the relatively low signal-to-noise-ratio (SNR) of ASL data. The aim of this study was to validate a new temporal denoising strategy for ASL MRI. Robust principal component analysis (rPCA) was used to decompose the ASL CBF image series into a low-rank component and a sparse component. The former captures the slowly fluctuating perfusion patterns while the latter represents spatially incoherent spiky variations and was discarded as noise. While there still lacks a way to determine the parameter for controlling the balance between the low-rankness and sparsity of the decomposition, we designed a method to solve this problem based on the unique data structures of ASL MRI. Method evaluations were performed with ASL CBF-based functional connectivity (FC) analysis and a sensorimotor functional ASL MRI study. The proposed method was compared with the component based noise correction method (CompCor). The proposed method markedly increased temporal signal-to-noise-ratio (TSNR) and sensitivity of ASL CBF images for FC analysis and task activation detection. We proposed a new temporal ASL CBF image denoising method, and showed its benefit for the CBF time series-based FC analysis and task activation detection.