Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.22901166The BGC-Argo datasets used in the proposed approach come from two distinct public repositories. First, BGC-Argo S-profile data from the Coriolis Argo Global Data Assembly Center (GDAC) (Argo, 2000; last visit in January 2026) were downloaded considering only delayed-mode (DM) and adjusted realtime (RT) data for the period from 1 July 2013 to 31 December 2024, ensuring a larger number of DM data. A quality-check procedure was applied as described in (Amadio et al., 2024), whose details can also be accessed via Zenodo. Specifically, the dataset was checked by retrieving only complete profiles with quality flags 1(good data), 2 (probably good data), 5 (value changed), and 8 (interpolated) for temperature, salinity,nitrate, oxygen, and chlorophyll. Additionally, a moving weighted average was applied to smooth out residual small-scale fluctuations, using the two upper and two lower neighboring points, with a Gaussian function of the vertical distance to the central point to determine the weights. The resulting curated dataset consists of approximately 229, 500 oxygen, 101, 000 chlorophyll, and 63, 000 nitrate profiles. The second BGC-Argo dataset, specifically for chlorophyll, was obtained from SEANOE (Schmechtig et al., 2025) (last visited in February 2026). It provides post-processed temperature, salinity and chlorophyll profiles from BGC-Argo incorporating a dedicated correction for optical measurements (Sauzede et al., 2025) resulting in chlorophyll estimates that are not directly equivalent to the GDAC Coriolis product. The same quality-control procedure applied to the Coriolis dataset was also applied to the SEANOE dataset. The GLOBIO models are trained and tested using nitrate data from Coriolis and chlorophyll data from SEANOE. In June 2026 (https://biogeochemical-argo.org/chla-reprocessing.php, accessed August 2026),the official GDAC adopted the chlorophyll reprocessing protocol proposed by SEANOE, initiating the replacement of the previously available chlorophyll data. For this reason, we present the model trained on the SEANOE dataset. As the CORIOLIS chlorophyll dataset remains available, we additionally trained a second model, using the CORIOLIS dataset. This additional experiment allows us to assess the robustness of the proposed approach to changes in the processing and quality control of chlorophyll profiles. The input data consist of both low-dimensional and vector-valued components: the former include temporal and geospatial variables, namely sampling year, day of the year, latitude, and longitude, while the latter correspond to vertical profiles of temperature, salinity, and oxygen measured by the float. Each original float profile is discretized onto a regular depth grid of N vertical levels. At each depth level, the profile value is estimated by averaging all observations within a neighborhood of half the discretization interval around the target depth. The resulting profiles are subsequently smoothed using a moving-average filter to reduce small-scale fluctuations while preserving their overall vertical structure. When no observations are available within the specified range, the value is obtained through linear interpolation between the two nearest available vertical levels. Temperature, salinity, and dissolved oxygen profiles are independently normalized using feature-wise z-score standardization, with the means and standard deviations computed over the training dataset. The temporal and geospatial variables are instead transformed to account for their specific characteristics. Latitude and longitude are converted to polar coordinates, while the day of the year is represented using its sine and cosine components. The sampling year is expressed as an offset from the minimum year in the dataset, i.e., year − minyear. No preprocessing is applied to the target nitrate profiles. Chlorophyll measurements exceeding 20, mg, m−3 are flagged as potentially unrealistic outliers and excluded from the dataset. Since the 1D CNN operates on vectorial inputs, the low-dimensional features are first projected into vector representations of length N . Three independent MLPs are employed for this purpose: one processes the two components of the day-of-year representation, one processes the year offset, and one processes the three polar-coordinate components. The resulting representations are concatenated with the three oceanographic profiles, yielding a six-channel tensor of shape 6 × N , which is then provided as input to the convolutional architecture. Each folder refers to a different dataset (nitrate, chlorophyll Seanoe and chlorophyll Coriolis). Each folder contains the corresponding train and test dataset in .csv
Keyword(s):
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)