There are three approaches to developing robust near infrared calibration models, including spectral pretreatment such as differentiation, Piecewise Multiplicative Scatter Correction (PMSC), Finite Impulse Response (FIR), and Orthogonal Signal Correction (OSC), to remove external variations, selecting wavelengths which are insensitive to external variations, and constructing temperature-hybrid calibration models. In this paper, these three strategies were investigated based on reforming gasoline NIR spectra collected at different temperatures in order to develop robust RON and benzene calibration models against temperature. It has been found that with only spectral pretreatment even OSC method fails to obtain satisfactory results, which could not remove the effects caused by temperature fluctuation. Selecting wavelengths by genetic algorithms and constructing temperature-hybrid calibration models, in which spectra measured at different temperature are combined into one calibration set, are both good approaches to developing robust NIR calibration models against temperature. The latter seems better because it needs no special knowledge and extra software, but thenon-linear effects should be considered in practical applications.