A data-driven approach to simulate the spatiotemporal variations of chlorophyll-a in Chesapeake Bay

Abstract

Phytoplankton biomass, indicated by chlorophyll-a (Chl-a) concentration, is fundamentally important for aquatic ecosystems. However, accurately simulating Chl-a is always challenging even when using state-of-the-art numerical models. We propose a data-driven modeling framework that combines Empirical Orthogonal Function (EOF) analysis and machine-learning technique to tackle this problem, using Chesapeake Bay as an example. Through the dimension reduction using EOF, the three-dimensional (3D) problem can be decomposed into multiple one-dimensional (1D) problems. The non-linearity of these 1D problems will be modeled with machine learning using an artificial neural network. Model performance in terms of spatiotemporal Chl-a variations with both seasonal and interannual signals is evaluated. The model performance is comparable or higher than 3D numerical models previously applied in Chesapeake Bay. Sensitivity tests reveal the necessity of forcing transformations to improve the model predictive skill. Instead of manually applying a transformation for each input forcing variable, an auto-selection procedure is adopted to choose an appropriate transformation from a variety of transformation options. While it is unlikely the data-approach can replace the traditional numerical models, we argue that data-driven approaches provide a promising way for future studies in coastal and estuarine systems considering the fast accumulation of observational data.