Release notes for version 2.01, September 2023:

• Adds new position, velocity, and SST-related variable estimates to previous version 2.00.
• This version adds data from drifters that were still alive at the cut-off date of v2.00 (30-Jun-2020 23:00:00) and data from recently released individual drifters that generated data after that date. This version now includes 197,214,787 estimates of position and velocity derived following the method described in Elipot et al. 2016 and SST (and related variables) from the methods described in Elipot et al. 2022. The new dataset spans 03-Oct-1987 13:00:00 to October 31 2022 23:00:00.
• Please note that because of the methods employed to process the data of each drifter trajectory, the variable estimates in v2.01 for the living drifters at the previous cut-off date will differ slightly from the estimates in v2.00.
• Some new variables have been added: start_date, start_lon, and start_lat that indicate the date and time and longitude and latitude of the first good observational data point as determined by the quality-control procedure of the GDP DAC at AOML. Note that these data differ from the deploy_date, deploy_lon, and deploy_lat data that correspond to the coordinates recorded at sea by the operator who deployed the drifter. In addition, the start_* variables do not necessarily correspond to the first estimated hourly data point for each drifter.
• An issue with the flagging of SST estimates was discovered in v2.00 and is now fixed in v2.01. As explained in Elipot et al. 2022, a quality flag 1 indicates that the estimation of the SST uncertainty variables has failed and therefore that the uncertainty variables (err_sst, err_sst1, err_sst2) should exhibit a _FillValue. Yet, in version 2.00, less than 0.18% of the SST estimates with quality flag 1 did exhibit finite uncertainty estimates and therefore were flagged incorrectly. Version 2.01 corrects this flagging error. The careful reader of Elipot et al. (2022) might ignore the part of the caption of Figure 7 where it is written that “The two populations […] are flagged with quality flag 1”. That is incorrect: these two populations can have any of the flags 3, 4, and 5.
• In Elipot et al. 2022, the flagging scheme for diurnal SST is based on the fulfillment of three criteria, denoted (1), (2), and (3). As written in the paper, a quality flag 2 should indicate that “none [of these criteria] are fulfilled”. This is incorrect and a quality flag 2 actually indicates that “the three criteria are not all fulfilled”. It other words, quality flag 5 indicates that criteria (1), (2), and (3) are fulfilled, quality flag 4 indicates that criteria (1) and (2) are fulfilled but not (3), quality flag 3 indicates that criteria (1) is fulfilled but neither are (2) and (3), and quality flag 2 indicates all other five combinations of fulfillment of the three criteria, including that none are fulfilled.

Release notes for version 2.00, September 2022:

• The hourly position and velocity dataset is now augmented by variables of sea surface temperature at hourly time steps: sea water temperature, non-diurnal sea water temperature, and diurnal sea water temperature, and their accompanying uncertainty estimates and quality flags. See Table 6 of Elipot et al. 2022.
• The position and velocity estimates are exactly the same as for version 1.04c (same number of drifters and same temporal range).
• The dataset is now officially distributed by the NOAA National Center for Environmental Information as a single NetCDF file containing contiguous ragged arrays, one for each data variable, as well as metadata.

Release notes for version 1.04:

• Version 1.04c: NetCDF files corrected to eliminate several errors including one-day offset, errors on GPS positions, and to clarify drogue metadata. Matlab files unchanged from 1.04(a).
• The interpolation methods for drifter trajectories tracked by the Argos system and by the Global Positioning Sytem are the same as for version 1.00, described in Elipot et al. (2016).
• An error was found in the calculation of GPS errors in version 1.01-1.02 which caused the errors to generally be overestimated. This was corrected for v1.03.
• The 1.04 update adds an additional 4 years and 9 months of quality-controlled data to v1.00, up to 4 April 2020. Starting in v1.02, we have added data before 1 September 2005 (the start date of v1.00) by considering all GPS-tracked drifters since the beginning of the GDP, and all Argos-tracked drifter data since the beginning of the GDP, when the average of the uneven sampling intervals per trajectory is less than or equal to 3 hours. As a consequence, the hourly dataset is a subset of the ' 6h dataset, which goes back to 1979.
• In v1.00, some drifters were erroneously categorized as being either Argos-tracked or GPS-tracked. This was corrected in v1.01. In addition, when drifters were tracked by both systems, only the GPS trajectory is contained in the dataset (starting v1.01). In v1.00, there were repeats when merging the Argos and GPS files.
• In verions 1.00 and 1.01, the WMLE (Argos) data were contained in a single, very large file. They have now been broken into smaller non-overlapping files, with file extensions "_blockN". Each file contains a unique set of drifters (no drifter is in multiple files). The only updated data for v1.03 is in block 7; blocks 1-6 contains drifters no longer alive as of v1.02, and will not be updated after v1.02.
• For version 1.04, the few Argos drifters that were alive as of v1.03 in block 7 were moved to a new block 8. Those that died before are retained in a smaller new block 7.
• The position and velocity estimates from GPS-tracked data are now provided (since v1.01) with formal 95% confidence intervals from the weighted least squares solutions of the estimation, assuming an observational error of 22 m for GPS positions. We diagnosed this observational error from the overall distribution of the GPS data received at the Data Assembly Center of the GDP. In order to obtain acceptable confidence intervals, the LOWESS processing of the GPS positions has been implemented to make sure that at least 3 data points are used for each estimation. This differs from v1.00 of the dataset for which only 2 data points were sometimes used.
• Simplified versions of the Matlab codes used to generate this dataset are available through a GitHub repository