Understanding and forecasting hypoxia using machine learning algorithms

Abstract

This study's primary objective lies in short-term forecasting of where and when hypoxia may transpire to enable observing its effects in real time, focusing on a case study in Corpus Christi Bay (Texas). Dissolved oxygen levels in this bay can be characterized by three temporal trends (daily, seasonal, and long-term). To predict hypoxic events, these three mathematical trends are isolated and extracted to obtain unbiased forecasts using a sequential normalization approach. Next, machine learning algorithms are constructed employing the continuous, normalized values from a variety of sensor locations. By including latitude and longitude coordinates as additional variables, a spatial depiction of hypoxic conditions can be illustrated effectively, allowing for more efficient summer data collection and more accurate, near-real-time projections. Using k-nearest neighbor and regression tree algorithms, approximate probabilities of observing hypoxia the following day were calculated, and estimates of dissolved oxygen levels were also computed. During periods in which hypoxia was observed, forecast probabilities of hypoxia exceeded 80%. Conversely, during periods in which no hypoxia was observed, the model's estimate remained below 20%. These results indicate that the modeling approach produces reasonable forecasts for this case study.