Classification-based event detection in ecological monitoring networks

Power-budgeting is a fundamental challenge in sensor networks today and the energy requirement of different sensing modalities is unevenly distributed. As a result, it is advisable to activate power-hungry sensors only during informative periods. Using low-power sensors, one can predict these informative periods due to strong correlations exhibited by environmental modalities. In this article, we consider an application of detecting “events” using classification based methods to increase the lifetime of the network. Specifically, we explore the problem of using low-power sensors to predict precipitation, which is one of the primary drivers of ecological activity. Such predictions can allow us to schedule the activation of expensive sensors (such as CO2) when they are most informative. In order to achieve this trade-off between power and collecting informative data, we focus our efforts on predicting/classifying precipitation based on features extracted from inexpensive ambient temperature and barometric pressure modalities. Experimental results obtained from weather data collected over multiple years demonstrates that we can achieve accuracy towards 80% using these low-cost modalities and simple linear classifiers. EJSE Special Issue: Wireless Sensor Networks and Practical Applications (2010) Figure 1 : An illustration of the typical signatures shown by Air temperature (AT) and Barometric pressure (BP) during the onset and departure of rain events. Note the drop in BP prior to the rain events and the deviation of AT from the diurnal pattern. and the Brier metric [12]. Experimental results based on 6 years of meteorological data show that the accuracy of using AT and BP features in conjunction with simple linear classifiers is towards 80% and are comparable to the performance of non-linear classifiers built on the same set of features. We find that the misclassification error of BP is lower than AT by as much as 15%. The brier score decomposition allows us to understand and analyze the probability of prediction. We find that quality of the prediction on both ends of the probability spectrum is fairly good and trustworthy. The rest of the paper is organized as follows. In Section 2 we introduce background information and survey similar bodies of work. In Section 3, we formulate the problem and provide a summary of the methods and metrics used to evaluate the solution. In Section 4, we present the dataset used for our study. The extraction of features is outlined in Section 5. In Section 6, we present our results, and finally, in Section 7, we conclude.