The breakup of the auroral substorm is closely related to a sudden electromagnetic energy release during solar wind-magnetosphere coupling process. Understanding the mechanisms of substorm onset and expansion phase clarifies the interactions among interplanetary magnetic field, magnetosphere and ionosphere. Additionally, substorm research is essential to characterize the process of flux transport from the solar to earth, which is significant to the space weather forecast. Auroral images from the ultraviolet imager (UVI) aboard the Polar satellite are the main dataset containing records of auroral substorms, with clear depiction of complete auroral ovals and substorm bulge features. Existing substorm detection algorithms are mostly empirical, relying on manually designed features and rules. In this article, we propose a detection algorithm guided by cascaded cluster algorithms for automatic substorm detection. To avoid using handcraft features, spatiotemporal features of UVI image sequences are extracted using a three-dimensional convolution network with subspace clustering. Because of imaging angles differences between frames, UVI images coordinates are converted into MLAT-MLT (the magnetic latitude-magnetic local time) coordinate system for pixel alignment. Moreover, image level clustering is applied to reduce the UVI image noise by isolating the substorm bulge and discarding unimaged areas. Experimental results indicate that the proposed method achieves higher recall than existing standard methods.