Spatio-temporal event detection using probabilistic graphical models (PGMs)

Event detection concerns identifying occurrence of interesting events which are meaningful and understandable. In dynamic fields, as time passes the attribute of phenomenon varies in spatial locations. Detecting events in dynamic fields requires an approach to deal with the highly granular data arriving in real time. This paper proposes a spatiotemporal event detection algorithm in dynamic fields which are monitored by wireless sensor networks (WSNs). The algorithm provides a method using probabilistic graphical models (PGMs) in WSNs to cope with the uncertainty of sensor readings. The algorithm incorporates the ability of Markov chains in temporal dependency modelling and Markov random fields theory to model the spatial dependency of sensors in a distributed fashion. Experimental evaluation of the proposed algorithm demonstrates that the decentralized approach improves the F1-score to 82% and 29% better precision than simple threshold technique. In addition, the performance of the algorithm was evaluated and compared with respect to the scalability (in terms of communication complexity). In comparison with the centralized approach the decentralized algorithm can substantially improve the scalability of communication in wireless sensor networks.