Assimilating a decade of hydrometeorological ship measurements across the North American Great Lakes

Abstract:

We use a decade of measurements made by the Volunteer Observing Ships (VOS) program on the North American Great Lakes to derive spatial estimates of over-lake air temperature, sea surface temperature, dewpoint, and wind speed. This Lagrangian data set, which annually comprises over 200,000 point observations from over 80,000 ship reports across a 244,000 square kilometer study area, is assimilated using a Gaussian Process machine learning algorithm. This algorithm classifies a model for each hydrometeorological variable using a combination of latitudes, longitudes, seasons of the year, as well as predictions made by the National Digital Forecast Database (NDFD) and Great Lakes Coastal Forecasting System (GLCFS) operational models. We show that our data-driven method significantly improves the spatial and temporal estimation of overlake hydrometeorological variables, while simultaneously providing uncertainty estimates that can be used to improve historical and future predictions on dense spatial and temporal scales. This method stands to improve the prediction of water levels on the Great Lakes, which comprise over 90% of America’s surface fresh water, and impact the lives of millions of people living in the basin.