As a key factor in the climate model, cloud phase is an important prerequisite to performing cloud property retrievals from remote sensor measurements. The ability to infer cloud phase using cloud emissivity spectra is investigated by numerical simulations. It is shown that for emissivity below 0.95, several spectral features such as the slopes, the ratios and the differences of the emissivity are consistent with the variation of cloud phase in some spectral regions. Specifically, these features include the slope of the cloud emissivity between 800 and 900 cm-1, the slope of the cloud emissivity between 900 and 1 000 cm-1, the difference in the mean emissivity between above-mentioned two regions, the ratio of the emissivity at 862.1 cm-1 to the emissivity at 989.8 cm-1, the difference in the emissivity between 862.1 and 989.8 cm-1, the ratio of the emissivity at 1 900.1 cm-1 to the emissivity at 2 029.3 cm-1, the ratio of the mean emissivity for far-infrared region to the emissivity at 900 cm-1. A cloud phase classifier is proposed based on support vector machines (SVM). A series of simulations including various cloud patterns are performed. The RBF kernel function parameters and the penalty factor of SVM are selected by using the genetic algorithm. The phase determination algorithm is applied for collecting data from the AERI at the SGP site. The results from the ground-based multisensor cloud phase classifier proposed by Shupe are used to validate the phase determination algorithm. It is found the two results are consistent in general. 30% clouds are indicated as opaque due to its high emissivity. The cloud with small lidar’s depolarization is misclassified as clear sky by the Shupe method. It can be concluded that the proposed algorithm considering the spectral information (spectral slopes, ratios and differences) is efficient for cloud phase determination of thin cloud.