In this letter, we propose a novel snapshot hyperspectral image segmentation system based on metasurfaces. This system combines a metasurface with a monochrome camera sensor chip to create a snapshot hyperspectral sensor, encoding the target hyperspectral image into an intensity pattern. A deep learning network is proposed for snapshot segmentation, which is capable of extracting spatial-spectral features of the measurement and directly predicting the segmentation results of the scene. The proposed system obviates the necessity of obtaining or reconstructing the hyperspectral image of the original scene. By doing so, it can not only simplify the optical system but also avoid the error that might occur during the hyperspectral image reconstruction. The results demonstrate that the proposed snapshot segmentation system achieves a prediction accuracy exceeding 99% in the context of land cover.