Event detection is an essential element for various sensor network applications, such as disaster alarm and object tracking. In this paper, we propose a novel approach to model and detect events of interest in sensor networks. Our approach models an event using the kind of spatio-temporal sensor data distribution it generates, and specifies such distribution as a number of regression models over spatial regions within the network coverage at discrete points in time. The event is detected by matching the modeled distribution with the real-time sensor data collected at a gateway. Because the construction of a regression model is computation-intensive, we utilize the temporal data correlation in a region as well as the spatial relationships of multiple regions to maintain the models over these regions incrementally. Our evaluation results based on both real-world and synthetic data sets demonstrate the effectiveness and efficiency of our approach.