Scientists from St. Petersburg State University have created a neural network model that analyzes the composition of water and allows determining the level of carbon dioxide in a body of water. This greatly facilitates monitoring the state of the ecosystem.
To measure carbon dioxide in water, oceanographers use the partial pressure (pCO ) indicator — the pressure that the gas would create if it alone occupied the entire volume. In oceanography, this parameter shows how saturated the water is with CO compared to the atmosphere. Specialists from St. Petersburg University analyzed data on environmental parameters affecting the concentration of carbon dioxide and built a model for assessing partial pressure using artificial intelligence.
In working with the neural network, scientists used data on the level of illumination, temperature indicators, and salinity of the water in the reservoir, as well as information about the depth of the mixed layer, obtained from open sources of expedition data SOCAT and from satellites.
The university reported that an excess of carbon in the water leads to acidification of water bodies. This destroys corals and shells. And when organic matter decomposes, the level of oxygen in the bottom layers decreases. This creates "dead zones." This is especially pronounced in the Baltic Sea.
Due to the large amount of river runoff and weak exchange with the ocean, cyanobacteria actively develop in the sea. They absorb a lot of carbon, which makes the water worse and there are more places where there is not enough oxygen. This threatens the ecosystem and makes it more vulnerable to climate change.
We have created partial pressure maps for the surface layer of the entire Baltic Sea, using real measurements of water parameters. Such maps allow for a more accurate assessment of the parameter in areas with rare measurements — for example, off the coasts of bays or in coastal zones in the autumn-winter period. Our model provides realistic indicators, confirmed by satellite and model data.
According to Sofya Kuzmina, a graduate of St. Petersburg State University, two types of data are used in machine learning: training and test data. Training data helps the model learn to determine which pCO values correspond to certain parameters. For example, when the water temperature decreases, the partial pressure may increase, since the solubility of the gas increases. The algorithm remembers this pattern and analyzes it in combination with other factors in order to effectively apply it in the future.
Read materials on the topic:
"Chemical Shepherds" will clean the sea of fuel oil: Russian scientists have created new reagents
RAS presented a biological method for cleaning the Black Sea from fuel oil
An unmanned boat for studying water bodies was created in St. Petersburg