The NPP processor comprises a set of Python-based routines that manage the ingestion, quality control, correction, and pre- and post-processing of autonomous underwater glider profiles as well as interfaces with external routines to calculate spectral PAR and NPP itself , which are implemented in the C programming language. Figure  shows a detailed flow diagram for the various processing stages. The processor supports multiple approaches to NPP calculation, depending on the availability of glider-based optical sensor data. Throughout this paper, when we refer to the NPP processor, we refer to the code routines that are represented in Fig. .

At its heart, the algorithms used to calculate NPP are as described in : these in turn draw heavily upon the spectral light NPP formulation of . However, this method is modified to cater for fluorescence quenching and light attenuation in shelf seas, as opposed to the open ocean (as discussed in Sect. ). For the purposes of determining NPP, the optimal glider instrument payload consisted of (in order of importance) chlorophyll-a fluorescence, PAR, and optical backscatter . Figure a shows that only four missions (497, 454, 499, and 517) had the full complement of required sensors, which necessitated modifications to the algorithms (see Table ).

The processing chain was designed to accommodate either near-real-time (NRT) mode or delayed-time mode (DM) implementations and as such ingests glider data either as individual netCDF profiles as they become available (in NRT) or in EGO netCDF format (in DM). The data set described here is processed in DM. Files may be ingested locally or auto-downloaded from a remote FTP repository on a user-determined schedule. All ingested source files are stored in an initial deployment directory and catalogued in a centralised SQLite database. This non-destructive approach supports the continual updating of the glider record from a remote catalogue in the NRT case while preventing replication. The database monitors, records, and manages all subsequent stages of the processor.

NPP calculations are performed on a profile-by-profile basis. Glider data typically consist of both downward (dive) and upward (climb) components in a single file, which in our processing framework represents two profiles. If a pre-existing profile designation is provided, as is usually the case in EGO data, this is used to split the source data into profiles. If no designation is provided, the ingested data are split into single files according to the turning points in the smoothed depth record. Smoothing is performed using a fifth-order Savitzky–Golay filter, with a nominal window of 51 points. This window, which represents 5–10 min in glider sampling time, does not relate to a particular physical scale but is short enough to allow smoothing to accurately capture the transitions between descending and ascending dive components but long enough to reduce incorrect dive splitting due to short inversions or “dwelling” at the top and bottom of dives. Individual profiles are then stored in the staging directory for future use.

Due to trade-offs necessary to achieve the multiple mission priorities of the AlterEco programme, not all gliders were able to accommodate PAR sensors. Here, when required, SEO-based PAR data are used in lieu of in situ measurements (a substitution that is covered in more detail from Sect.  onward). This increases both the flexibility and utility of the method for the operational oceanography community, allowing it to be applied to glider data where only in situ chlorophyll-a fluorescence is available. In addition, SEO and reanalysis data are used to provide information on the prevailing atmospheric and marine conditions during each glider mission. A list of SEO and reanalysis data sources and the variables that are extracted and/or derived can be found in Table  and in Fig. .

When a glider mission is updated (e.g. a new profile is added in NRT mode), the processor calculates the new temporal and spatial extents of the mission. Using these extents, the processor gathers the required SEO and reanalysis data from the specified source, concatenates the retrieved catalogue in time, and trims the spatial coverage to produce a “data cube” that matches the glider mission extents. The spatial trimming is performed remotely, on the server side, if the data service in use allows this capability, reducing data transfer costs and time. In NRT mode, new data are added to extend the cube as required, without the need to download the entire catalogue once again (e.g. via the concatenation of new time slices to the existing local record). This operation is performed for all variables and for all gliders, irrespective of whether they have the relevant in situ measurement, allowing for the continual validation of the use of SEO and reanalysis data as a substitute.

Once the data cube has been constructed, the average time and location of each profile are extracted and concatenated into a one-dimensional time series of the glider trajectory. Bi-linear interpolation is then used to retrieve the corresponding SEO and reanalysis data from the relevant data cube, resulting in an SEO trajectory file for each variable, with a value for each profile. During construction, the cube is both spatially and temporally “padded” to eliminate “edge effects” associated with interpolation.

The pre-processing step consolidates the glider and SEO-based data on a profile-by-profile basis, performs quality control procedures, and selects the relevant variables for NPP calculations depending on availability (Fig. ). Sporadic missing values are common in in situ data. Where possible, linear interpolation is used to fill these gaps in the positional, depth, and pressure data. If interpolation is not possible, the profile is discarded, and no further processing takes place.

Following this, and where not provided directly, conservative temperature and absolute salinity are calculated from the glider CTD record using the TEOS-10 Python GSW toolbox

https://teos-10.github.io/GSW-Python/ (last access: 2 June 2022)

Once the pre-processing stages have been completed for all available glider profiles (Fig. ), spectral Ed profiles are calculated for each glider profile using the solar irradiance model described by . To account for local meteorological conditions, the model runs using the total column atmospheric ozone ([O3]), cloud cover, wind speed, relative humidity, and total column water vapour parameters for each profile are stored in the relevant trajectory file (Table ). These spectral Ed values calculated from the model are scaled such that their integrated value between 400 and 700 nm matches the corresponding Eo+ measurements provided by the glider or SEO data sources (see Sect. ). This scaling correction accounts for instantaneous sky conditions associated with each profile.

The work by implemented a novel methodology to exploit the relationship between PAR and [Chl-a] to account for changes in the apparent fluorescence to chlorophyll calibration, brought about by phytoplankton community succession. This approach allows dynamic changes to the calibration and reduces the need for in-field calibration, which is difficult, if not impossible to implement, especially in the near-real-time case. However, this method is based on an in-water model suitable for Case 1 waters , with the non-water component of light attenuation ascribed to [Chl-a] only .

The processor retains the ability to implement this method, as detailed extensively in and represented in Fig.  by the “Case 1” decision box. However, the optically complex waters of the shelf seas are rich in sediment and do not conform to the Case 1 paradigm. Implementation of a spectral irradiance model more suitable to the region requires the consistent deployment of in situ PAR and backscatter sensors that is not available across the AlterEco programme. Consequently, no PAR-based scaling of the [Chl-a] profiles is performed for this data set. This caveat is further discussed in Sect. .

Net primary production, P, is calculated from the corrected [Chl-a] and spectral downwelling PAR profiles using the model, as presented in and shown in Eq. (). The model calculates NPP through a triple integral across day length (L), depth (D1=0, D2=Zeu), and wavelength (λ1=400 nm, λ2=700 nm). The absorption cross section per unit of chlorophyll (a*; m2 g-1]) and net growth rate (ϕμ; mol(carbon) mol(quanta)−1]) are parameterised as in . 4P=12gmol-1d-1∫0L∫D1D2∫λ1λ2[Chl-a](Z)Ed(t,Z,λ)a*(λ)ϕμ(t,Z,λ)dλdZdt NPP estimates, in units of carbon flux (mg m-2 d-1), are calculated for all corrected [Chl-a] profiles using the per-profile average time and position for each. The piecewise measurements are integrated from the 1 % light level (as determined by the model) to the surface to give a final estimate of depth-integrated primary productivity in carbon flux of milligrams per square metre per day.

Figure  allows a comparison of using SEO-based PAR in the calculation of spectral Ed and subsequent NPP in contrast to using in situ PAR. SEO-based PAR is shown to function as a suitable proxy in this method, remaining highly correlated with its in situ counterpart, with mean values that are within 2 % of the target estimate.

When combined, the NPP times series derived from the AlterEco glider deployments spans a 19-month period, as shown in Fig. . As expected, NPP is at its greatest in the spring and early summer and reaches its highest in the spring of 2019, corresponding with the timing of the regional spring bloom. Conversely, it drops to near zero in the winter months, when light availability becomes limiting. The figure also shows the inherent spatial and temporal variability in the time series, reflected in the inter-glider and intra-glider data, respectively. Despite operating during the same period, glider 454 (Cabot) measures approximately twice the NPP of glider 455 (Orca) throughout April, May, and June 2018, where we expect biological activity to be near its highest level.

Black traces, indicating NPP estimates derived from in situ PAR sensors, compare well with their coloured (SEO-derived) counterparts in all cases. However, the divergence between the signals recorded by the concurrently deployed gliders 455 (Orca), 497 (Humpback), and 454 (Cabot) strongly suggests the presence of significant variability in the region north of the Dogger Bank (Fig. b, lower-right corner).

In order to quantify the efficacy of the various quenching correction algorithms upon the chlorophyll fluorescence profiles, a comparison was made between the daytime (quenched) and night-time (not quenched) [Chl-a] over the top 20 m of the water column. This comparison assumes that there is little change in the vertical profile of [Chl-a] in a 24 h period: this is obviously a simplification as there will be changes due to (1) bloom growth or loss (e.g. via respiration, decay, and grazing) and (2) spatial variability. Figure  shows that the major discrepancy between the night-time and the uncorrected daytime [Chl-a] profiles for glider 454 (Cabot; see Fig.  and Table ) is when the integrated [Chl-a] in the top 20 m exceeds 4 mg m-2 and is less than ∼12 mg m-2. The percentage variance between daytime and night-time uncorrected cases is 91.0 % with a significant p value of <0.01. The method (Fig. ) clearly outperforms the other correction methods tested in this case (r2=97.8%, p<0.01; cf. : r2=32.8%, p<0.01; cf. : r2=0.2%, p=0.81). Similar results were obtained for glider 517 (Cabot; see Fig.  and Table ), with (r2=97.8%, p<0.01) performing best, followed by (r2=73.3%, p<0.01) and (r2=26.2%, p<0.01).

Figure  shows the corrected [Chl-a] profiles for glider 454 (Cabot). The corrected time series shows a gradual deepening of the DCM over the 3-month mission from around 25 m in mid-May (2018) to 40 m in mid-August. During this period there is also a clear reduction in the peak [Chl-a]: >3 mg m-3 at the start of the mission, ∼1 mg m-3 towards the end. The transects on and off Dogger Bank (depths shoaling to around 40 m) are clearly correlated with a shoaling of the euphotic depths from ∼30 m on the bank to ∼50 m off the bank. Changes in the MLD are less pronounced, apart from a deepening between 18 and 22 June 2018 and a rapid shoaling from 30 to 18 m around 24 June 2018. Meteorologically, June 2018 was characterised by a relatively slack pressure gradient (light winds) until a brief 3 d period of stronger north or north-westerly winds on 19 and 22 June, which corresponds to the episodic deepening of the MLD in this period.

Two distinct advantages of gliders are that they sample flexibly, in terms of horizontal space and depth, and they can gather data at high frequency. As there are no pre-existing measurements of NPP in the North Sea with comparable frequency, here we compare our results with available estimates of annual mean productivity from both satellite and in situ sources.

Independent [Chl-a] estimates, derived from v4.2 of the ESA Ocean Colour Climate Change Initiative (OC-CCI) data set , indicate that the 2018 spring bloom was relatively intense in comparison to 2019, a fact that does not appear to be reflected in the glider NPP data. applied the NPP methodology of and to this satellite [Chl-a] record, producing a 20-year time series of satellite-based NPP from 1998–2018. Although this record does not span the entire AlterEco period (2017–2019), is only calculated at 9 km resolution, and is only available as a monthly product, it provides a useful data source to compare the glider NPP measurements against. Satellite-based NPP estimates for each glider are shown in light orange in Fig. . Comparing the glider and co-located satellite time series further suggests that 494 (Stella) likely failed to fully capture the onset of the spring bloom in 2018. However, given the disparity between the estimates obtained by 454 (Cabot), 494 (Stella), and 497 (Humpback), it is likely that the region is subject to significant spatial heterogeneity, which perhaps the satellite product is too coarse to record.

Table  summarises the monthly and annual NPP estimates across all AlterEco glider campaigns. The monthly mean and standard deviations derived from the satellite NPP record, calculated across a box spanning the AlterEco sampling region, are shown in the final two columns of Table 5. The annual cycle and mean annual NPP rate as measured from the glider missions agree well with contemporaneous values interpolated from the monthly mean OC-CCI NPP record. It is notable from the table that, while still within 1 standard deviation, the April NPP peak in the glider data is somewhat lower than its OC-CCI counterpart, likely due to the low signal recorded by 494 (Stella) over this period in 2019. However, the glider-based NPP signal peaks at a time consistent with remote sensing estimates.

The glider-based annual mean NPP value is 98 g C m-2 a-1. This compares favourably with the 119 g C m-2 a-1 measured by , who applied a 14C approach to measure daily NPP through extensive surveys carried out over ICES Region 7 (north of Dogger Bank), scaling up to monthly estimates using the mean daily value across the region and number of days per month. It also compares well to the annual estimate of 125 g C m-2 a-1 for the northern North Sea proposed by , based on a synthesis of daily NPP estimates from multiple cruises. The glider measurements are similarly consistent with NPP estimates derived from models, with recording 130 g C m-2 a-1 for ICES Region 7 (as used by ), simulating 119 g C m-2 a-1 across the northern North Sea, and reporting 82.6–118.8 g C m-2 a-1 for the central and northern North Sea in their tidal simulations.

Alongside NPP, provide estimates of gross primary production (GPP). In the northern North Sea (ICES Region 4), an NPP of 149 g C m-2 a-1 is associated with a GPP of 314 g C m-2 a-1. Assuming that the ratio of NPP : GPP remains broadly constant in the region on an annual basis, we can apply this to our glider NPP measurements to obtain a GPP estimate of an approximate value of ∼200 g C m-2 a-1. This compares favourably with the measurements of , who reported an annual mean gross production of 200±15 g C m-2 a-1 in seasonally stratified regions from 1998 to 2013 (including Dogger Bank).

Primary production is highly variable on short temporal and spatial scales. The impact of the mesoscale variability associated with fronts and eddies can be extensive. High-frequency changes in tidal phase , sky conditions, and the local wave field can also exert a strong influence. To monitor the impact of these processes in highly productive shelf seas, it is desirable to continually sample key regions using technologies that support adaptive sampling strategies. Autonomous underwater vehicles (AUVs), such as gliders, offer one such approach to this problem, offering persistent monitoring of shelf sea biogeochemistry and informing regional model assimilation strategies .

This data set presents the first intra-annual, glider-based in situ NPP time series for the North Sea that is able to address questions pertaining to biophysical interactions on a high-frequency basis. From Fig.  it is clear that the NPP signal is modulated at multiple frequencies within individual deployments, and substantial spatial heterogeneity exists between co-deployments (e.g 455 (Orca) and 454 (Cabot)). Further analysis of this data set should give insight into the physical processes that contribute to this variability.

When deployed with multiple mission goals in mind, glider payload space typically comes at a premium. Most notable in this case is the effect on the deployment of PAR sensors, which are present on less than 50 % of missions. However, adaptation of previous methodology to accommodate SEO-based PAR estimates has been shown to be feasible. Combining SEO surface data with AUV profiles also presents interesting options for reconstructing subsurface fields. Machine learning methods have demonstrated the feasibility of combining SEO surface fields with in situ profiles to render a three-dimensional picture of ocean biogeochemical properties . The data set presented here would be well suited for application of such methods to evaluate and further extend coverage of NPP data in the global ocean.

The processing method developed here allows for glider-based NPP to be calculated in a much broader array of cases. While in DM it can replicate the approach of , the inclusion of differing quenching algorithms promotes application to different regions and/or different sensor loads (e.g. those without backscatter). The flexible inclusion of SEO data in lieu of in situ PAR measurements expands this utility even further, allowing NPP calculations from gliders with a more limited array of sensors without substantial loss of accuracy (Fig. ). Finally, the ability to support NRT ingestion of glider data allows for NPP calculation in an operational setting.

PAR, when spectrally decomposed, can be used to provide a calibration of the [Chl-a] fluorometer . Although the fluorometer calibration may be accurate at the start of an individual mission, calibration using nearby discrete [Chl-a] samples at launch and retrieval of the glider may lead to a false sense of security, particularly in areas of high heterogeneity, such as experienced during this study. showed that within-mission variability in the correction factor is possible due to changes in phytoplankton community structure. However, the model previously proposed is suitable for Case 1 waters only and does not account for absorption and scattering by CDOM and sediment, respectively, and so no dynamic calibration is applied here. The strong agreement between AlterEco glider NPP measurements and both satellite and historical in situ estimates (see Sect. ) underlines the validity of the data set, and future work will consider the incorporation of a model to cater for more complex waters, where glider payload allows.

As noted in Sect. , the measurement of in situ dark counts for fluorescence is performed on the entire glider mission. This method is therefore inappropriate for near-real-time analysis of glider profiles. Inclusion of a methodology to calculate dark counts for both the fluorescence and backscatter measurements on a per-profile basis, such as that developed by , would also be advantageous.

While the quenching correction method of proved most appropriate in this case, this result should not be considered a general solution. This rationale underpins the decision to incorporate multiple methods to correct near-surface fluorescence; however, the method eventually chosen is limited by the sensors deployed and most notably the availability of backscatter data (Fig. ). The availability of backscatter data allows for a wider selection of correction methodologies in both DM processing and NRT processing . In addition, its inclusion is essential to constructing a complex water model, as discussed earlier in this section. As the NPP processor was developed during the AlterEco programme, which commenced in 2017, it only takes advantage of quenching methods available at the time. Future work is expected to include more recent quenching methodologies such as .

For long-duration missions (i.e. more than a few days) bio-fouling of sensors mounted on gliders can affect data quality. Unlike Argo floats, which typically park at depths well below the euphotic zone (∼1 km), for 10 d, gliders spend a greater portion of their time in the photic zone, allowing the build-up of a bacterial substrate and then algal colonisation. Despite many strategies to mitigate bio-fouling (copper-covered sensors, bio-wipers), it is impossible to completely eradicate it currently, and even predicting its onset is problematic. Anecdotally on moorings situated in the western English Channel , bio-fouling has been observed to take several months to colonise sensors and then following cleaning has only taken a few weeks to re-emerge. Best efforts have been made to truncate the glider [Chl-a] record where bio-fouling appears evident (Sect. ).

Here, the methodology described is used to generate a primary productivity data set in an optically complex shelf region. However, much of the basis of the methodology is derived from previous work that was developed for use in the open-ocean context (e.g. ). Consequently, we expect the NPP processor to be viable in the open ocean, where chlorophyll concentration tends to dominate the optical signal. In the open ocean, quenching methods based on calibration against the backscatter record are also likely to perform better (e.g. ) and, in the case of , allow for dynamic calibration associated with changes in phytoplankton community structure. As Earth-observation-based retrieval of chlorophyll concentration typically has lower errors in the open ocean, there may be opportunities to investigate the use of remotely sensed data to correct and dynamically calibrate the in situ chlorophyll record, an approach previously suggested by . It is, however, important to point out that the methodology may require tuning when used in different mission contexts. With deeper and/or longer dives, care should be given to select the correct smoothing parameters to determine the turning points of the profile. In addition, where in situ PAR is not available, it may also be advisable to select a KdPAR‾ model that is more suited to clear waters, such as . More broadly, future investigations should also consider the effect that the choice of KdPAR‾ model used has on the resulting NPP value.