Two-phase CNN for model data fusion: Predicting 3D chlorophyll-a in the Mediterranean Sea

Chlorophyll-a is a key indicator of marine ecosystem state and variability, monitored through satellite observations, in-situ measurements, and estimates from process-based models. Each source, however, suffers from intrinsic limitations such as incomplete coverage, multiple-source uncertainties, or simplified model representation. Reconstructing chlorophyll-a fields that accurately captures ecosystem variability, therefore, requires effective integration of these heterogeneous data through model–data fusion, which remains a major challenge. We introduce MuSt3Net, a deep-learning framework based on a Convolutional Neural Network (CNN) that performs model–data fusion through a sequential two-step learning strategy to predict the 3D distribution of chlorophyll-a in the Mediterranean Sea. MuSt3Net generates 3D chlorophyll-a fields by (i) learning the relationships between physical drivers and chlorophyll-a as produced by a process-based model, and (ii) integrating sparse in-situ BGC-Argo observations to propagate their information across the full 3D domain. This structured training scheme enables the network to preserve typical spatial patterns learned from the process-based model while integrating local corrections informed by observations. The reconstructed fields capture the characteristic spatial and seasonal variability of the Mediterranean Sea. Validation against independent BGC-Argo profiles confirms the model’s skill, yielding a mean RMSE of 0.08 mg m−3. Winter surface blooms are reproduced with an error of approximately 0.1 mg m−3, and the deep chlorophyll-a maximum is represented with a vertical error below 10 m. These results demonstrate the effectiveness of the proposed two-step fusion strategy for generating improved 3D biogeochemical reconstructions.