The phase-sensitive time-domain reflectometer ($\phi \text{-}\mathrm{OTDR}$) has been popularly used for events detection over a long period of time. In this study, the events classification methods based on convolutional neural networks (CNNs) with different features, i.e., the temporal-spatial features and time-frequency features, are compared and analyzed comprehensively in $\phi \text{-}\mathrm{OTDR}$. The developed CNNs aim at distinguishing three typical events: wind blowing, knocking, and background noise. The classification accuracy based on temporal-spatial images is higher than that based on time-frequency images (99.49% versus 98.23%). The work here sets a meaningful reference for feature extraction and application in the pattern recognition of $\phi \text{-}\mathrm{OTDR}$.