A Supervised Learning Approach to Water Quality Parameter Prediction and Fault Detection

Water quality parameters such as dissolved oxygen and turbidity play a key role in policy decisions regarding the maintenance and use of the nation's major bodies of water. In particular, the United States Geological Survey (USGS) maintains a massive suite of sensors throughout the nation's waterways that are used to inform such decisions, with all data made available to the public. However, the corresponding measurements are regularly corrupted due to sensor faults, fouling, and decalibration, and hence USGS scientists are forced to spend costly time and resources manually examining data to look for anomalies. We present a method of automatically detecting such events using supervised machine learning. We first present an extensive study of which water quality parameters can be reliably predicted, using support vector machines and gradient boosting algorithms for regression. We then show that the trained predictors can be used to automatically detect sensor decalibration, providing a system that could be easily deployed by the USGS to reduce the resources needed to maintain data fidelity.