Spatio-Temporal Signal Recovery Based on Low Rank and Differential Smoothness

The analysis of spatio-temporal signals plays an important role in various fields including sociology, climatology, and environmental studies, etc. Due to the abrupt breakdown of the sensors in the sensor network, there always are missing entries in the observed spatio-temporal signals. In this paper, we study the problem of recovering spatio-temporal signals from partially known entries. Based on both the global and local correlated property of spatio-temporal signals, we propose a low rank and differential smoothness based recovery method (LRDS), which novelly introduces the differential smooth prior of time-varying graph signals to the field of spatio-temporal signal analysis. The performance of the proposed method is analyzed theoretically. Considering the case where a priori information about the signal's global pattern is available, we propose prior LRDS to further improve the reconstruction accuracy. Such improvement is also verified by synthetic experiments. Besides, experiments on several real-world datasets demonstrate the improvement on recovery accuracy of the proposed LRDS over the state-of-the-art spatio-temporal signal recovery methods.