Sea temperature monitoring from air temperature data, a neural networks approach

This thesis investigates the feasibility of reconstructing sea temperature using air temperature data from coastal weather stations. Neural networks have been used for their capability of modelling non linear relationship. The study focuses on the Northern Adriatic Sea, utilizing data from weather stations in Ancona, Venezia, and Trieste to reconstruct SST at different temporal and spatial scales. The investigation comprises three main experiments: (1) reconstruction of point measurements from the Acqua Alta oceanographic tower using various temporal processing approaches (hourly, filtered hourly, and daily data), (2) reconstruction of spatially averaged temperatures from CMEMS reanalysis data with different vertical levels and horizontal domains, and (3) pointwise reconstruction of the full temperature field. While the analysis revealed limitations in reconstructing Acqua Alta measured temperature, further investigated through coherence analysis, more promising results were achieved in reconstructing both spatially averaged and pointwise temperature fields from the CMEMS reanalysis. The results were analyzed through spatial and temporal distributions of performance metrics, revealing systematic patterns in reconstruction accuracy. A comparative analysis between satellite observed temperature, reanalysis and neural networks reconstruction provides insights into the relative strengths and differences of these three approaches to temperature estimation.