Machine Learning Approaches for the Quality Control of Tide Gauge Observations

Generating high-quality, continuous time series of 6-minute water level observations from raw tide gauge data is notoriously challenging. Data can be impacted by sensor, communication or other problems due to a range of issues such as extreme events, aging instrumentation and accidents. Though some problematic or spurious data are removed during initial automated quality-control (QC) steps, questionable water level values often remain in the records prior to human investigation. Once problematic data points are removed from the primary water level sensor time series, considerable effort and care are still required to fill the resultant data gaps with observations from the back-up sensor, nearby neighbor sensors, tide predictions, statistical fit of data around the gap or a combination thereof. The NOAA National Ocean Service (NOS) Center for Operational Oceanographic Products and Services (CO-OPS) operates and maintains over 250 real-time water level stations across the coastal U.S. and Great Lakes. This equates to over 1.8 million 6-minute water level observations per month which are processed, quality controlled and reviewed. CO-OPS presently relies on a combination of automated and manual processes, however a substantial amount of human intervention is required to generate a complete, verified water level time series. Here we present an initial investigation into the application of Artificial Intelligence (AI) approaches to process, QC and fill water level observations. A range of Machine Learning (ML) techniques are explored, including regression, random forests and artificial neural networks to assess performance and skill compared to target data sets of manually verified water level time series. Initial performance is assessed for over 10 years of 6-minute observations from five NOAA tide gauges (approximately 4.4 million records). An AI water level processing system has the potential to substantially reduce the work-hours required to produce high-quality water level time series, while improving the lag between data collection and verification to near real-time.