The gravitational pump plays a significant role in the carbon cycle by exporting sinking organic carbon from the surface to the deep ocean. The deep sediment time-series traps provide unique observations of this sequestered carbon fluxes. However, the sinking particles trapped are sensitive to physical short-term spatial and temporal fluctuations, which present a challenge in establishing a connection with their surface origin. In this study, we present a novel machine learning tool, designated as U-Net_sst-ssh, which is capable of predicting the catchment area of particles trapped at the PAP (Porcupine Abyssal Plain) station STs moored at a depth of 3000 m, based solely on satellite data. The model was trained and evaluated using Lagrangian experiments in a realistic CROCO numerical simulation. In comparison to the conventional approach, which yielded an average prediction of only 20-30% of the source area, the U-Net_sst-ssh predictions enhanced this score to 40-60%.

The dataset comprises U-Net_sst-ssh predictions that were applied to satellite observations of sea surface temperature (SST) and sea surface height (SSH) at PAP. This resulted in the creation of a catchment area dataset spanning 22 years. The catchment areas are calculated as a two-dimensional density probability function for a presumed collection period of 10 days and at a 10-day interval. The temporal variable of the dataset corresponds to the mid-date of the particle surface origin. The catchment areas are contingent upon the particle sinking rate (w), with five distinct sinking speeds (80, 100, 150, 200, and 300 m/d) being considered.

The dataset was validated by demonstrating a stronger correlation between the deep fluxes measured at PAP and the surface chlorophyll, compared with traditional catchment area methods. However, it was shown that the precision of the predictions remain highly sensitive to the local deep dynamics.