Hyperspectral technology is a rapid, nondestructive and accurate means of mineral detection, which can clearly reflect mineral chemical composition change. Garnets have the characteristic of three diagnostic peaks in the thermal infrared wavebands, and the wavelength positions of the reflection peaks are closely related to the chemical composition, so the subclass classification of garnets can be studied according to the thermal infrared wave spectrum characteristics. Reflection peak wavelength positions of uvarovite and spessartine are easy to distinguish with other types. However, that of almandine and pyrope, andradite and grossular have a large overlap and are difficult to distinguish with each other. Therefore, a fast and accurate classification method based on the thermal infrared spectrum is urgently needed. In this paper, the information about wavelength position and difference between wavelengths of the three reflection peaks of 85 different types of garnet samples were obtained from the thermal infrared spectroscopy library. Three nonlinear BP neural network methods, cluster analysis and multiple linear discrimination analysis were used to carry out garnet subclass recognition experiments, and the accuracy rate, recall rate and F1 value were used to evaluate the classification accuracy. The experimental results showed that the accuracy rate, recall rate and F1 value of BP neural network algorithm after classification could all reach 100%, and all types of garnets got a good distinction; the accuracy rate, recall rate and F1 value of clustering analysis and multivariate linear discriminant analysis were 86.1%, 80%, 79.2% and 84.2%, 80%, 79.5% separately, the four types of garnets with overlapping reflection peaks could not be well differentiated. According to the results, the nonlinear BP neural network is more suitable for the subclassification of garnets. Our study used powerful automatic nonlinear mapping ability of the BP neural network, has found the complex mapping relationship between the wavelength positions of the reflection peak in the thermal infrared spectrum of the garnets and the subclass types, and proved the feasibility and superiority of BP neural network method combined with thermal infrared spectrum characteristics. The identification of garnet subclass provided is fast and effective, and it can give good technical enlightenment for the rapid and effective identification of other minerals.