Improving maximum daily salinity regressor performance in the Columbia River estuary project

Sensors deployed in the Columbia River Estuary

As part of an Environmental Observation and Forecasting System, sensors deployed in the Columbia River Estuary collect salinity, temperature, and pressure data. This data collection is vital to the CORIE project members, who are constantly working to have a better understanding of the river dynamics. Of these sensors, salinity sensors are subject to bio-fouling, an event caused by growth of biological material in the sensor. This degrades the performance of the sensor over time, causing an approximately linear decrease in the maximum daily salinity measured by the instrument. Traditionally, the CORIE staff would look at time series, determine when bio-fouling started, and disregard all data from that point.

Clean and bio-fouled salinity time series examples from sensors in the Columbia River estuary. The top timeseries shows a clean signal, with its natural variability. The bottom plot shows a bio-fouled sequence. Notice how the maximum salinity value is decreasing over time after 9/28, the estimated time at which bio-fouling began. Image from Archer et al[1].

Archer et al[1]. proposed an automatic fault detection mechanism, which used a single Gaussian regressor to predict the maximum daily salinity using temperature information. This allowed them to look at growing diverging trends (i.e., difference between observed and expected increases), thus alerting the CORIE staff of a biofouling event. This reduced significantly the data loss.

Prof. Todd K. Leen invited me to this project as part of my Master's thesis. We improved the performance of the regressor by looking at temperature measurements as inputs. We modelled the joint density of salinity and temperature measurements as a Gaussian Mixture Model. The expected value of the mixture is a nonlinear regression function.

Here is my final thesis draft.

References

[1] Cynthia Archer, Todd K. Leen, and António Baptista. Parameterized Novelty Detection for Environmental Sensor Monitoring. Advances in Neural Information Processing Systems 16: 619-626, 2004.

For more information on the CORIE project, visit their website.