The intermittency of atmospheric turbulence plays an important role in the understanding of particle dispersal in the atmospheric boundary layer and in the statistical simulation of high-frequency wind speed in various applications. There are two kinds of intermittency, namely the magnitude intermittency (MI) related to non-Gaussianity and the less studied clusterization intermittency (CI) related to long-term correlation. In this paper, we use a 20 Hz ultrasonic data set lasting for one month to study CI of turbulent velocity fluctuations at different scales. Basing on the analysis of return time distribution of telegraphic approximation series, we propose to use the shape parameter of the Weibull distribution to measure CI. Observations of this parameter show that contrary to MI, CI tends to weaken as the scale increases. Besides, significant diurnal variations, showing that CI tends to strengthen at daytime (under unstable conditions) and weaken at nighttime (under stable conditions), are found at different observation heights. In the convective boundary layer, the mixed-layer similarity is found to scale CI exponent better than the Monin-Obukhov similarity. At night, CI is found to vary more slightly with height in the regime with large mean wind speeds than in the regime with small mean wind speeds, according to the Hockey-Stick theory.