The database presented in this study is available from PANGAEA, Data Publisher for Earth and Environmental Science in two formats: (1) the database containing all compiled raw fluorescence profiles (the raw database, http://doi.pangaea.de/10.1594/PANGAEA.844212, Sauzède et al., 2015b) and (2) the database containing the fluorescence profiles which are calibrated into chlorophyll a concentration and associated phytoplankton community size indices (the calibrated database, http://doi.pangaea.de/10.1594/PANGAEA.844485, Sauzède et al., 2015c). The data of in situ vertical fluorescence profiles compiled for creating the raw database were obtained from several available online databases as well as published and unpublished individual sources. The duplicates and single-surface values, which are not vertical profiles, were automatically removed (not integrated in the raw database). Finally, the raw database contains 268 127 fluorescence profiles. Following a robust quality control procedure detailed hereafter (Sect. 2.2), about 49 000 chlorophyll fluorescence profiles were converted into phytoplankton biomass (i.e., chlorophyll a concentration) and size-based community composition (i.e., microphytoplankton, nanophytoplankton and picophytoplankton). The origin of this calibrated database is summarized in Table 1. The majority of the data come from the National Oceanographic Data Center (NODC) and the fluorescence profiles acquired by Bio-Argo floats are available on the Oceanographic Autonomous Observations (OAO) web platform (63.7 and 12.5 % respectively, see percentages of data in the database depending on their origin in Table 1).

Different modes of acquisition were used to collect the data presented in this study: (1) the CTD (conductivity, temperature and depth) profiles are acquired using a fluorometer mounted on a CTD rosette; (2) the OSD (Ocean Station Data) profiles are derived from water samples analyzed by fluorometry and are defined as “low” resolution profiles (Boyer et al., 2009); (3) the UOR (Undulating Oceanographic Recorder) profiles are acquired by a “fish” equipped with fluorometer and towed by a research vessel; (4) AP (Autonomous Platforms) profiles are acquired by Bio-Argo profiling floats or elephant seals equipped with a fluorometer (Claustre et al., 2010b; Guinet et al., 2013). Table 2 lists the number of profiles in the calibrated database according to these four modes of acquisition.

It is worth noting that the data acquired from gliders were not included in the database. Although glider data are extremely numerous, they are restricted to a very small spatio-temporal window. As a consequence, a database including glider data would likely be spatially and temporally biased, in contradiction with our first aim of building a global climatological database.

In order to use the FLAVOR method (see details in Sect. 2.3), a specific and adapted data quality control procedure was developed and applied to each in situ chlorophyll fluorescence profile. This procedure was schematically implemented according to four main steps of data control (Fig. 1), each step being developed for discarding most, if not all, spurious fluorescence profiles that would deteriorate the quality of the database. Firstly, several basic tests were applied: (1) duplicates and single-surface values, which are not vertical profiles, were removed (these profiles were removed from the beginning of the process so they are not included in the so-called raw database); (2) coastal profiles were removed using a bathymetric mask of 500 m depth; (3) the uppermost measurement has to be located within the 0–10 m layer, while the deepest measurement has to be at or below 100 m. Secondly, tests on the profile vertical resolution are applied: (4) a minimum of 10 values per profile is required (i.e., condition on the vertical resolution acquisition); (5) a minimum of five non-equal values per profile are required (i.e., condition on the sensor resolution). Then, several tests are applied on the fluorescence profile shape. These conditions are based on the parameter used for the development of the FLAVOR method, Z0, which is the depth at which the fluorescence profile returns to a constant background value (see details in Sect. 2.3 and examples in Fig. 1b and c). (6) The median of the fluorescence values from the surface down to 20 m has to be greater than the median of the values of the last 10 % of the deepest samples of the profile (see Fig. 1a); (7) the depth Z0 has not to be within the last 10 % of the deepest samples of the profile (see Fig. 1b). Finally, a test on the noise of the profiles was developed and applied: (8) profiles with aberrant data caused by electronic noise are removed (i.e., variability greater than 20 % of the total profile range, see Fig. 1c). To finish, a visual check allowed all the remaining fluorescence profiles to be verified. The number of raw fluorescence profiles rejected at each step of the quality control procedure is presented in Table 3. Around 80 % of the raw fluorescence profiles were thus removed by this procedure. This step is an essential prerequisite for the development of a “clean” database of vertical distributions of phytoplankton biomass and community composition in the global ocean. The quality control procedure removed 77, 71, 28 and 25 % of the OSD, UOR, AP and CTD profiles, respectively, with profiles removed by the test on the bathymetry not taken into account.

In order to assess the vertical distribution of the total chlorophyll a concentration (hereafter, [TChl]) and the chlorophyll a concentration associated to each phytoplankton size index (hereafter, [microChl], [nanoChl] and [picoChl] for microphytoplankton, nanophytoplankton and picophytoplankton respectively), the FLAVOR method (Sauzède et al., 2015a) is applied to each chlorophyll fluorescence profile, satisfying the quality control procedure (see Sect. 2.2). In summary, FLAVOR is a neural network-based method which uses (1) the shape of the chlorophyll fluorescence profile (10 values from the normalized profile with values range between 0 and 1); (2) the depth Z0, which is the depth at which the fluorescence profile returns to a constant background value (see examples of Z0 depths represented by the horizontal red line for two profiles on Fig. 1b and c); and (3) the location (latitude and longitude) and the day of acquisition of the fluorescence profile as inputs. The outputs of FLAVOR are the vertical distributions of (1) [TChl] and (2) [microChl], [nanoChl] and [picoChl] with the same vertical resolution as the input raw fluorescence profile. FLAVOR is composed of two different neural networks: the first one was adapted to retrieve the vertical distribution of [TChl] and the second one to retrieve the vertical distributions of [microChl], [nanoChl] and [picoChl] simultaneously. Both neural networks were adapted and validated using a large database including 896 concomitant in situ vertical profiles of HPLC pigments and chlorophyll fluorescence. These profiles were collected as part of 22 oceanographic cruises representative of the global ocean in terms of trophic and oceanographic conditions, making the method applicable to most oceanic waters. The diagnostic pigment-based approach of Uitz et al. (2006), based on Claustre (1994) and Vidussi et al. (2001), was utilized to estimate the biomass associated with the three pigment-derived size classes for each profile. Finally, the data set of concurrent fluorescence profiles and HPLC-determined [TChl], [microChl], [nanoChl] and [picoChl] at discrete depths was used to establish the neural network-based relationships between the fluorescence profile shape and the vertical distributions of [TChl] and phytoplankton community. The schematic overview of the FLAVOR method is shown on Fig. 4 in Sauzède et al. (2015a). The global absolute errors of FLAVOR retrievals are 40, 46, 35 and 40 % for the [TChl], [microChl], [nanoChl] and [picoChl], respectively (Sauzède et al., 2015a).

Admittedly, the FLAVOR method has some limitations. The dependence of chlorophyll fluorescence on the light environment is probably intrinsically accounted for in the algorithm thanks to the geolocation and date of acquisition used as inputs for the training. However, one of the potential concerns with FLAVOR is that the impact of the daytime non-photochemical quenching (NPQ; see, e.g., Cullen and Lewis, 1995), responsible for a decrease in chlorophyll fluorescence values at high irradiance, is not accounted for by the method. The NPQ uncorrected fluorescence profile shape is indeed used to retrieve the vertical distribution of phytoplankton biomass (see details in Sauzède et al., 2015a). Note that, if density profiles are available together with fluorescence profiles, NPQ can be corrected using the method of Xing et al. (2012). This method involves substituting the fluorescence values acquired within the mixed layer by the maximum value within this layer.

It has been previously mentioned that FLAVOR is not adapted for the retrieval of chlorophyll a concentration on a fluorescence profile-by-profile basis (Sauzède et al., 2015a). Rather, FLAVOR and, hence, the resulting database, are relevant for large-scale investigations, e.g., development of climatologies of the vertical distribution of chlorophyll a, from which regional anomalies or temporal trends might be evidenced. In fact, the method was validated using a global database and it is not excluded that the retrievals from FLAVOR might be regionally biased. For instance, Sauzède et al. (2015a) have shown that FLAVOR retrievals for the Southern, Arctic and Indian oceans are slightly less accurate than for the other basins. This is likely because the method is not constrained enough in these specific areas which are known for data scarcity. Additional details about the performance of the method for various oceanic basins are given in Sauzède et al. (2015a), in Figs. S3, S5–S7. Finally, it is worth recalling here that the relationships between the phytoplankton biomass or community composition profiles and the fluorescence profiles are assumed to be identical for profiles acquired before 1991 (not involved in the training data set because of lack of HPLC data) and after 1991 (only used for the training process). In the context of possible use of this database for supporting analysis in looking for trends or a shift in chlorophyll a time series, this assumption will have to be taken into consideration.

An additional step of quality control is further applied once the FLAVOR method has been operated. It is based on Chauvenet's criterion which is used to identify statistical outliers in the retrieved biomass data (Buitenhuis et al., 2013; Glover et al., 2011; O'Brien et al., 2013). The criterion was applied to the surface data of each profile (median of values from the surface down to 20 m). As Chauvenet's criterion is based on the assumption that the data follow a normal distribution, the analysis was performed on the log-normalized [TChl] surface values. Such a criterion removes aberrant data partially caused by the failure of the FLAVOR method (see number of profiles removed by Chauvenet's criterion in Table 3).

The 48 600 chlorophyll fluorescence profiles which successfully passed all the steps of quality control were transformed into total chlorophyll a concentration and associated phytoplankton community size indices (i.e., microphytoplankton, nanophytoplankton and picophytoplankton) using FLAVOR (see details in Sect. 2.3). The resulting database covers all ocean basins with more profiles in the Northern Hemisphere (75 %) than in the Southern Hemisphere (25 %, see Figs. 2 and 3). However, the Southern Hemisphere remains relatively well represented with the profiles acquired by autonomous platforms and especially by elephant seals equipped with a fluorometer. Few data were acquired in the Indian Ocean and in the tropical South Atlantic and South Pacific (see Fig. 2). The highest numbers of fluorescence profiles are found at the BATS (the Bermuda Atlantic Time-series Study) and HOT (the Hawaii Ocean Time-series) time-series stations, which are located at 31.67∘ N–64.17∘ W and 22.75∘ N–158.00∘ W, respectively, and where data acquisition started in 1988. On the annual scale, the data acquisition appears evenly distributed, with a slight underrepresentation of autumn months (April to June) in the Southern Hemisphere (Fig. 4). The temporal distribution of fluorescence profiles in the database covers 56 years from 1958 to the present (Fig. 5a) and most of the observations were collected after the late 1980s. There are fewer observations from 2010 to 2012 because all data generally acquired by ship-based platforms have not been archived yet in the online databases. A significant increase in data density observed between 2013 and 2015 (in 2015, 124 profiles were acquired in half a month) mainly results from data acquired by Bio-Argo profiling floats. Around one-sixth of this global database has been sampled in only 2 years by the Bio-Argo platforms. This illustrates the potential of this new type of acquisition which is expected to dramatically increase the number of collected fluorescence profiles in the future.

Vertically, the database includes values of total chlorophyll a concentration and associated phytoplankton community composition from the surface down to a mean sampling depth of 743 m (with a maximum sampling depth ranging from 100 to 6000 m; Fig. 5b).

We present the database with respect to the vertical distribution of the total chlorophyll a concentration ([TChl]). Figure 6 displays the median [TChl] gridded within squares of 3∘ latitude by 3∘ longitude and over six vertical layers (0–25, 25–50, 50–75, 75–100, 100–150 and 150–200 m). In the surface layer (0–25 m, see Fig. 6a), the [TChl] median is the highest in the North Atlantic and the lowest in the South Pacific subtropical gyre. The median [TChl] decreases with depth for all the data, except for data acquired in South Pacific and Atlantic subtropical gyres where the median [TChl] increases with depth. This increase is associated with the so-called deep chlorophyll maximum (DCM) that is typical of these oligotrophic regions (e.g., Cullen, 1982; Mignot et al., 2011, 2014).

The global distribution of the phytoplankton community composition, given in terms of fraction of chlorophyll a concentration associated to micro-, nano- and picophytoplankton, is presented for the 0–1.5 Ze layer (Fig. 7a, b and c respectively). Here Ze, the euphotic depth is defined as the depth at which the irradiance is reduced to 1 % of its surface value. It was estimated according to the method of Morel and Berthon (1989), using the [TChl] profiles derived from FLAVOR. Figure 7 reveals general geographic patterns which are consistent with the knowledge about the ecological domains and biogeochemical provinces (Longhurst, 2010). On average microphytoplankton are dominant in the subarctic zone, with a relative contribution to the chlorophyll biomass reaching more than 70 % in these areas (Fig. 7a). Picophytoplankton are dominant in the subtropical gyres (South and North Pacific as well as South and North Atlantic), with a contribution reaching 45–55 % (Fig. 7c). Nanophytoplankton appear to be ubiquitous with a relatively stable contribution to biomass of 40–50 % (Fig. 7b).

To further assess the quality, range and representation of the FLAVOR-retrieved [TChl] database presented in this study, the retrieved surface [TChl] is compared to the remotely sensed [TChl]. In this context, the climatological [TChl] mean was extracted at a 9 km spatial resolution from NASA Modis Aqua archive for the time period covering 2002 to 2014. The extracted satellite [TChl] data were re-gridded to a 3∘×3∘ spatial resolution. Similarly the FLAVOR-retrieved [TChl] values for the upper layer of the database (i.e., mean value calculated between the surface and 20 m) for the same period were re-gridded to 3∘×3∘ squares. Figures 8 and 9 show that climatological averaged [TChl] from Modis Aqua and from the present database are generally consistent (Fig. 8a and b). The log-transformed ratio of the Modis Aqua to the database [TChl] estimates reveals a rather good agreement with a median value of -0.16 and a standard deviation of 0.58 (see histogram in Fig. 8c). Figure 9 displays the geographic distribution of the log-transformed ratio between the Modis Aqua and the database estimates of climatological surface [TChl]. The ratio shows no specific spatial bias. However, as it is mentioned in Sect. 2.3, FLAVOR retrievals for the Southern, Arctic and Indian oceans are slightly less accurate than for the other basins; it is therefore possible that the estimation errors are greater in these areas. Moreover, this observation has to be nuanced considering the difficulties in retrieving accurate ocean color satellite [TChl] in these high-latitude environments (Gregg and Casey, 2004; Guinet et al., 2013; Johnson et al., 2013; Peloquin et al., 2013; Siegel et al., 2005; Szeto et al., 2011).

As an example of application, monthly climatologies were computed for three ecological provinces defined by Longhurst (2010) and well represented in the current data set (Fig. 10a): (1) the North Atlantic Subtropical Gyral Province West (NASW, Fig. 10b), (2) the Atlantic Subarctic Province (SARC, Fig. 10c) and (3) the North Pacific Subtropical Gyre Province (NPTG, Fig. 10d). Overall the time series of the vertical distribution in [TChl] are consistent with expectations as detailed by Longhurst (2010). For the NASW province (Fig. 10b), [TChl] is relatively homogeneous from the surface to around 140 m from January to March; then the stratification of the water column leads to the establishment of a deep chlorophyll maximum (DCM) from April to November. Over the year, [TChl] varies in a restricted range of values (0.35–0.55 mgm-3). The dominant phytoplankton groups are the nano- and the picophytoplankton with relative chlorophyll contribution reaching 40–45 % for both size-based groups. The contribution of microphytoplankton remains low throughout the year (10 %). For the SARC province, the phytoplankton bloom starts in May (as indicated by a significant increase in [TChl], Fig. 10c). The bloom continues for 4 to 5 months with [TChl] within the 1.5–2 mgm-3 range (with maximum values in July). The microphytoplankton contribution increases during the bloom and reaches a maximum (60 %) in August, whereas the nanophytoplankton relative contribution decreases from April to August. The contribution of picophytoplankton increases slightly all along the year to reach a maximum of about 40 % in December. For the NPTG province (Fig. 10d), a DCM (0.15–0.25 mgm-3) is established at a depth of 100–125 m and persists all year long. This DCM deepens in summer, consistently with a deeper light penetration in the water column at this period. The [TChl] at DCM reaches a maximum value in June and July. The dominant phytoplankton groups are the nano- and the picophytoplankton with relative contribution reaching 45–50 % for both size-based groups and slight opposite temporal evolutions. The contribution of microphytoplankton remains low throughout the year (<10 %).