Mixture density networks are a better predictor of nutrients than chlorophyll in a well studied, optically complex estuary

It has long been recognized that empirical algorithms that predict water quality from satellite remote sensing have poor accuracy in optically complex coastal, estuarine and inland waters. This is unfortunate as it these environments that are most susceptible to anthropogenic stressors and therefore stand to benefit the most from the promise of highly resolved monitoring from space. To fill this critical void, machine learning approaches have emerged as a superior method to predict water quality measurements in optically complex environments. In this study we begin by assembling a large dataset of multispectral Sentinel 3 data collocated in space and time with the data rich Chesapeake Bay water quality monitoring program. Following two common methods, we compare and contrast the segregation of data across Optical Water Types (OWTs via K-means clustering), and assess the efficacy of Mixture Density Networks (MDNs) to predict in-situ water quality data. Critically, we extend our analysis beyond traditional optically active constituents (e.g. chlorophyll, Secchi depth, total suspended solids) and include bulk total nitrogen and phosphorus concentrations. OWTs vary predictably along the estuarine salinity gradient and generally yield statistically different distributions of water quality parameters across spectrally dissimilar OWTs. Perhaps surprisingly, MDN predictions of total nitrogen and phosphorus not only have small errors (32% and 38% respectively), but they also exceed the measurement error of chlorophyll (72%).

Presentation Preference: Oral

Primary Presenter: Greg Silsbe, University of Maryland Center for Enviornmental Science (gsilsbe@umces.edu)

Authors:

Anna Windle, NASA  (anna.windledipaola@nasa.gov)

Sairah Malkin, University of Maryland Center for Env. Science  (smalkin@umces.edu)

Raleigh Hood, University of Maryland Center for Environmental Science  (rhood@umces.edu)