Abstract. Marine chlorophyll a concentration is a key indicator of ecosystem health, and accurate seasonal forecasting has important applications for fisheries management, harmful algal bloom detection, and climate studies. Traditional approaches rely on computationally intensive numerical biogeochemical models that require extensive domain expertise and parameterization. We propose a data-driven, resource-efficient alternative: a neural architecture based on the U-Net that reconstructs surface, near-global chlorophyll a based on four physical predictors. The model learns to emulate satellite-like chlorophyll a from mixed layer depth, sea surface height, salinity, and temperature, all of which are known to influence phytoplankton distribution and nutrient availability. By leveraging publicly available seasonal forecasts of these variables, we can generate six-month of chlorophyll a predictions in a matter of minutes with a single GPU. We trained the model using the GLORYS12 reanalysis as input physics and GlobColour merged chlorophyll a observations as reference. When applied to seasonal forecasting by using SEAS5 ensemble forecasts as input as opposed to the reanalysis, the model maintains high skill globally and remains stable across the six-month forecast horizon. Regional analysis shows that the model accurately captures seasonal dynamics and bloom timings across diverse regimes, with performance comparable to or exceeding that of a state-of-the-art numerical biogeochemical model, while requiring orders of magnitude less computational resources. This approach demonstrates that high-performing, seasonal chlorophyll a forecasting can be achieved through a resource-efficient, observation-driven framework, offering a practical alternative for operational applications where computational constraints limit the use of full biogeochemical models.
Balbontin et al. (Fri,) studied this question.