Float Data
Synthetic profile files were downloaded from the Argo Global Data Assembly Center (2023–05 snapshot, doi:10.17882/42182). There were 631 floats equipped with Seabird Scientific chlorophyll fluorescence sensors from 30 March 2008 to 09 May 2023 (Fig. 1a), with 503 ECO sensors, and 128 MCOMS sensors. All floats carried conductivity-temperature-depth sensors, and a subset of floats (n=234) carried downwelling radiometer sensors (Seabird Scientific OCR-504) that included measured radiance at 490 nm (Fig. 1b). Profiling float data go through post-deployment quality control following standard Argo protocols16,38. For ChlFL data, the field CHLA_ADJUSTED represents ChlFL data that has gone through the QC process, and the CHLA field represents raw ChlFL based on applying factory calibration coefficients in mg m-3. For ECO and MCOMS sensors, the dark counts are subtracted from the raw sensor counts, and then converted to CHLA in mg m-3 by applying a factory-determined scale factor. The scale factor converts fluorescence to Chl-a concentration based on a single calibration with a monospecific culture of the diatom Thalassiosira weisflogii. During the QC process, four main adjustments are applied to CHLA16: determination of in situ dark counts based on deep measurements, non-photochemical quenching correction during the day, manual inspection and flagging of bad and questionable data, and a bias correction of two13. The CHLA_ADJUSTED field should have these adjustments applied, but there are small inconsistencies in the details of how these adjustments are implemented, particularly for the in situ dark count and NPQ corrections. Therefore, to eliminate this source of uncertainty, we have applied the dark and NPQ correction39 using the CHLA field for all of the floats and omitted the global bias correction of 2. However, to take advantage of the manual inspections that flagged bad data, data quality flags were imported from CHLA_ADJUSTED of 1, 5, and 8, which corresponds to good data, value changed, and estimated value. The latter two flags are applied for NPQ corrected and interpolated data.
Briefly, for each float, an in situ dark correction was applied by subtracting the median of the minimum ChlFL value of the first five deep (> 900m) profiles from all data17. Floats that were recently deployed and did not collect at least five deep profiles were not included. Daytime profiles (defined as having a sun angle > 0 using MATLAB function SolarAzElq) were adjusted for non-photochemical quenching39. This correction finds the maximum ChlFL value above the mixed layer depth (defined as a density change greater than 0.03 kg m-3 from a surface reference value40), and copies that value from its coinciding depth to surface. ChlFL values > 50 mg m-3 and less than 0.014 mg m-3 were removed from the dataset. 50 mg m-3 is a reasonable upper limit for open ocean chlorophyll-a maxima, whereas the lower limit is twice the factory-specified sensitivity of 0.007 mg m-3. This limit of detection was confirmed in situ by looking at the smallest change between samples in the mixed layer depth of night-time profiles of floats (WMO ID’s 5906514, 5904655, 5906529, 5904172) in a low chlorophyll region near Hawaii. To compare float ChlFL to ChlSAT, the median value of ChlFL was calculated over the first optical depth (OD) because this is roughly equivalent to the depth of ocean color satellite retrievals. The first optical depth was estimated per profile as the inverse of Kd(490), estimated using ChlSAT using the following equations (Morel et al.20, Eq. 8),
Kd(490) = 0.0166 + 0.077298 x (ChlSAT)^0.67155
OD = 1/Kd(490)
Float temperature and salinity data were used to calculate the mixed layer depth. For this, adjusted temperature, pressure, and salinity data with Argo quality flags of 1, 2, 5, and 8 were used when available. If only un-adjusted data were available, quality-control flags of 1, 2, 3, 5, and 8 were used. Quality control flags of 1, 2, 3, 5, and 8 were used for downwelling irradiance data. A visual inspection of the irradiance data was used in Ocean Data View to remove floats and profiles with obviously bad irradiance data.
ChlKd
ChlKd, the estimate of Chl-a concentration based on the attenuation of light, was estimated from radiometric measurements of light irradiance at 490 nm19, using the subset of floats carrying both a radiometer and fluorometer. Only the irradiance data to the first optical depth were used, rather than a threshold depth of minimum light, in order to be consistent with ChlSAT and the median surface ChlFL used for the study. We found that setting the integration depth to first optical depth versus the mixed layer depth affected the final bias correction values (Supplementary Fig. 7), suggesting that this is an important definition for similar analyses. Only profiles with a sun angle greater than 10 degrees above the horizon were used to estimate ChlKd. A 7-point median filter was applied to each profile to minimize the effects of wave focusing at the surface, passing clouds, or changes in the float’s position with respect to vertical. The attenuation coefficient was determined by Model 1 regression slope of depth versus the natural log of irradiance down to the first optical depth. ChlKd was estimated by inverting Equation 8 from Morel et al.20 to solve for chlorophyll-a. This chlorophyll-a estimate represents a water column average between the surface and the first optical depth. Profiles with Kd(490) less than 0.0166 (the attenuation due to water) were excluded and profiles with an R2 fit > 80%, and a relative standard deviation of the estimated slope < 10% were used to estimate ChlKd.
Satellite Data
Ocean color products derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard the NASA Aqua satellite were used for this analysis. The level-3, 8-day averaged, 9 km resolution ChlSAT concentration product (OCI algorithm21,22) was downloaded from the NASA Ocean Biology Processing Group. The 8-day averaged product was chosen to improve spatial coverage of the satellite data, which can be limited in daily satellite observations due to incomplete global satellite coverage, high sun glint, clouds, as well as the infrequency of same-day float-satellite matchups. ChlSAT data less than 0.05 mg m-3 were removed based on the minimum value of in situ Chl-a data used in algorithm development21,22.
For each float profile, ChlSAT data that were within 8 km and closest in time based on the median satellite date were matched. The 8 km threshold was chosen based on an autocorrelation threshold analysis over one year of 4 km (highest level-3 spatial resolution) MODIS Chl (01 Jan 2020 to 31 Dec 2020). Globally, this 8 km threshold corresponds to 75% or higher autocorrelation for 99.97% and 96.59% of valid matchups across longitude and latitude, respectively (Supplementary Fig. 8). The autocorrelation threshold analysis was completed separately for both the zonal and meridional directions based on similar methods41. Briefly, each line of latitude or longitude was treated as a discrete data series, for which an autocorrelation function can be calculated (done using MATLAB function autocorr), and defining a length scale at lag m≥0.75. This lag indicates the first location within the data series at which the resulting autocorrelation coefficient is less than or equal to 0.75. The median of all spatiotemporally matched satellite data per profile was used to compare to the float data.
Building the climatology
The ratio of ChlFL data to the median of ChlSAT matchups or ChlKd were taken per profile, and represent the final data used to estimate climatological correction factors, where ChlFL and ChlKd are median values taken within the first optical depth. Within a 5°×5° gridded area, the median of ChlFL:ChlSAT or ChlFL:ChlKd data for a single month, year and float are taken first, then the median of this data across floats for a month and year, and finally the median and standard deviation across all years for a month is taken, producing the final gridded climatological median and standard deviation values presented here. Seasonal climatological values are taken as the median of monthly climatological values for the northern/southern hemispheres, respectively, for December, January, February (winter/summer), March, April, May (spring/fall), June, July, August (summer/winter), September, October, November (fall/spring). Amplitudes were calculated as the absolute difference between the maximum and minimum monthly climatological values for each 5°×5° grid with more than 6 months of valid data, and the negative inverse of bias corrections < 1 was taken prior to calculating amplitudes. Area-weighted values are reported for calculated global medians.
The climatology presented here includes both daytime and nighttime float data. To limit potential errors introduced by the NPQ correction, it would be preferable to use only night-time profiles of ChlFL, however this would have greatly reduced our number of valid ChlFL profiles by approximately 75%, and removed the comparison to ChlKd, which is only valid during daylight hours. A linear trend between the annual gridded climatology using night-only and daytime-only profiles showed good agreement with minimal additional bias (R2 = 0.6, m = 0.9) (Supplementary Fig. 6). To illuminate potential regional discrepancies between day and night-time data, the difference between the two were taken for the annual climatological mapped data (Supplementary Fig. 6). In general, regions showed no consistent or unique differences when using one data set versus the other, with the exception of the Eastern Equatorial Pacific where bias corrections would be lower if using daytime data compared to night time data.