Combining machine learning with data assimilation to improve the quality of phytoplankton forecasting in a shelf sea environment

TL;DR

This study combines machine learning with data assimilation to significantly enhance short-term phytoplankton forecasts in the North-West European Shelf, addressing biases caused by excess nitrate concentrations.

Contribution

It introduces a neural network model to predict surface nitrate and integrates it into operational forecasting, improving phytoplankton forecast accuracy by up to 30%.

Findings

Nitrate assimilation improves forecast skill by up to 30%.

Flow-dependent nitrate data outperform climatology-based approaches.

Hybrid ML and data assimilation methods are feasible for operational use.

Abstract

We demonstrate that combining machine learning with data assimilation leads to a major improvement in phytoplankton short-range (1-5 day) forecasts for the North-West European Shelf (NWES) seas. We show that excess nitrate concentrations are a major reason behind known biases in phytoplankton forecasts during late Spring and Summer, which can grow fast with lead time. Assimilating observations of nitrate would potentially help address this, but NWES nitrate data are typically not available in sufficient abundance to be effectively assimilated. We have therefore used a recently developed and validated neural network (NN) model predicting surface nitrate concentrations from a range of observable variables and implemented its assimilation within a research and development version of the Met Office's NWES operational forecasting system. As a result of nitrate assimilation the phytoplankton forecast skill improves by up to 30%. We show that although much of this improvement can be achieved by using a weekly nitrate climatology predicted by the NN model, there is a clear advantage in using flow-dependent nitrate data. We discuss the impacts of this improvement on a range of additional eutrophication indicators, such as dissolved inorganic phosphorus and sea bottom oxygen. We argue that it should be feasible to implement this hybrid machine learning - data assimilation approach within the near-real time NWES operational forecasting system.