Dottorato in TECNOLOGIE DELL'INFORMAZIONE E DELLE COMUNICAZIONI (ICT)

Titolo della tesi: Integration of satellite and in situ data to measure CO2 fluxes in the Mediterranean Sea

Although carbon dioxide stands as a primary driver of future climate, the carbon cycle lacks complete characterization, particularly in the ocean. Satellite measurements of pCO2 are possible through measurable proxies. Given the complexity of carbon cycle processes, the use of regional optimization has often proven successful. Despite the Mediterranean Sea's climate sensitivity, insitu measurements of carbon-related quantities are sparse and there is a lack of literature-acknowledged regional satellite algorithms to estimate marine pCO2. This study introduces the first year and a half of measurements made at Lampedusa of ocean pCO2 and computed CO2 fluxes, which represent the first data available in the Central Mediterranean regarding flux calculation. The net effect, over one whole year, is an absorption of (2.23 +/- 0.04)x10-2 kg(CO2)/m2, and preliminary analysis suggests the impact of the 2022-2023 marine heatwave leads to a substantial reduction of the ocean CO2 absorption. Gaps are present and more in-depth analyses are needed. Regional-optimized algorithms for pCO2 estimation using satellite-derived quantities are introduced, employing both traditional and machine learning multiple regression approaches. The best traditional model, which estimates pCO2 as a function of SST, CHL and PAR, exhibits a bias of 4.4 μatm, RMSD of 13.0 μatm, and a coefficient of determination of 0.94. The leading machine learning model, which relates pCO2 to SST, achieves a bias of -1.4 μatm, RMSD of 25.0 μatm, and an R2 of 0.76. The limited dataset size poses a challenge, particularly in machine learning model training. The models were also used to compute fluxes. Both approaches demonstrated good agreement, with R2 values of 0.75 for the traditional approach and 0.64 for the machine learning approach. The biases were -0.081x10-9 kg m-2 s-1 and -0.30x10-9 kg m-2 s-1, and RMSD values were 1.3x10-9 kg m-2 s-1 and 1.5x10-9 kg m-2 s-1 for the traditional and machine learning approaches, respectively. While a satellite-based approach shows promise, there is room for improvement, including the use of a broader dataset for further advancements.