Deep learning applied to the analysis of dissolved carbon dioxide in coastal areas of the Balearic Sea

This work studies the changes in the sea surface dissolved carbon dioxide (CO2 (diss.)) at the Estación de Investigación Costera del Faro de Cap Salines, a pristine site at the Southeastern tip of Mallorca (GPS 3.05457°E, 39.26552°N). More specifically, the dependence of CO2 (diss.) on a number of recorded parameters like atmospheric carbon dioxide (CO2 (atm.)) and temperature rise, which may be the result of direct or increased human activities, is uncovered by the modelling of CO2 (diss.). This modelling of CO2 (diss.), is done with Deep Learning (DL) methods applied to a relatively small ecological dataset and then these DL methods are compared to each other. This work also shows and generalises how DL methods can give reasonable results when applied to small tabular datasets in general. Along with the analysis of different DL methods (Deep Symbolic Regression, Permutation Convolutions and Self-Attention), an architecture based on different methods (Self-Attention, Mish activation function, Ranger optimiser and Learning rate finder), which are selected after reviewing and testing ideas from the literature, is also presented in this work.

Supervisors:  Manuel A. Matias and Iris Eline Hendriks

Jury members: Emilio Hernández-García, Miguel C. Soriano, and Manuel A. Matias

The defense will presential, broadcasted through Zoom:

https://zoom.us/j/98286706234?pwd=bm1JUFVYcTJkaVl1VU55L0FiWDRIUT09