Dealing with missing sensor values in predicting shellfish farm closure

Shellfish farms need to be closed from harvesting when the water body is contaminated to avoid a serious health hazard. We have designed a sensor network framework to monitor a number of water quality variables to check the health of shellfish farms and predict closure if hazardous. Because of the uncertainty associated with the data acquisition process, a full set of sensor values are not always available for decision making purposes. The prediction system thus needs to deal with missing values. Statistical approaches are commonly used to generate an artificial value to approximate a missing sensor reading and predictions are made on the then complete set of sensor values. In this paper we present a new method that is capable of making predictions without making artificial approximations of missing values. The idea is to train a set of classifiers on different subsets of sensor values. Given a set of available sensor values, a prediction is made by the classifier trained on the corresponding set of sensor values. We have evaluated the system on the data obtained from a number of shellfish farms in Tasmania. Experimental results demonstrate that the proposed method to deal with missing values can predict closures with high accuracy.