GLOBIO: Bridging Global and Local Scales for Biogeochemical Profiles Prediction

Studying the state of marine ecosystems, their changes over time, and the influence of human activities requires accurate ocean observations. However, reliable measurements are sparse and unevenly distributed across both time and space, with significant disparities in coverage among different variables. In recent years, autonomous oceanographic instruments such as Argo profiling floats have improved the collection of subsurface data. Despite this advancement, physical variables like temperature and salinity, and oxygen are more affordable to monitor, while sensors for biogeochemical variables—such as nitrate, chlorophyll, and bbp700—remain costly. This disparity results in a gap between the abundance between physical and biogeochemical measurements, confirming the need for methods that estimate biogeochemical variables using high-frequency physical data to fully leverage ocean observing systems like Argo.

Some existing ANN-based techniques, which rely on Multilayer Perceptron (MLP) architectures trained on point-wise ship-based measurements, allow for the prediction of, e.g., nitrate profiles by exploiting intrinsic information contained in the input profiles of T, S, and oxygen. Alternatively, we propose an approach that directly infers the vertical profile in a single step using a spatially aware neural network.

Using a spatial-aware neural network, we propose an approach that directly infers the entire vertical profile in a single step. By leveraging the typical shape of biogeochemical profiles as a learnable constraint, the model can fully exploit the potential of the BGC-Argo dataset.

A regional approach using a spatial-aware neural network has already been proposed in “PPCon 1.0: Biogeochemical Argo Profile Prediction with 1D Convolutional Networks”. However, PPCon was limited to the Mediterranean Sea, while our objective is to develop a global-scale model.

Given PPCon’s promising results—demonstrating smooth and accurate profile predictions with improvements over previous MLP applications—we extend this approach by developing a global 1D CNN to predict nitrate, chlorophyll, and backscattering (bbp700) from date, geolocation, and profiles of temperature, salinity, and oxygen.

PPCon’s promising results demonstrated smooth and accurate profile predictions in the Mediterranean Sea, showing improvements over previous MLP applications, particularly for chlorophyll and bbp700, while nitrate performance remained comparable. 

We build on this approach by developing a global 1D CNN using a quality-checked dataset of 101,000 chlorophyll-a and 63,000 nitrate profiles spanning 2012 to 2024. 

Additionally, the new approach incorporates transfer learning, enabling a pre-trained model to be fine-tuned on different datasets by replacing and retraining the final layers of the network. 

The novel 2-step method and preliminary results will be presented to highlight the potential for the automatic generation of regional models.