Towards a multi-platform assimilative system for ocean biogeochemistry

Oceanography has entered an era of new observing platforms, such as biogeochemical Argo floats and gliders, some of which will provide three-dimensional maps of essential ecosystem variables on the North-West European (NWE) Shelf. In a foreseeable future operational centres will use multi-platform assimilation to integrate those valuable data into ecosystem reanalyses and forecast systems. Here we address some important questions related to glider biogeochemical data assimilation and introduce multi-platform data assimilation in a (pre)operational model of the NWE Shelf-sea ecosystem. We test the impact of the different multi-platform system components (glider vs satellite, physical vs biogeochemical) on the simulated biogeochemical variables. To characterize the model performance we focus on the period around the phytoplankton spring bloom, since the bloom is a major ecosystem driver on the NWE Shelf. We found that the timing and magnitude of the phytoplankton bloom is insensitive to the physical data assimilation, which is explained in the study. To correct the simulated phytoplankton bloom one needs to assimilate chlorophyll observations from glider or satellite Ocean Color (OC) into the model. Although outperformed by the glider chlorophyll assimilation, we show that OC assimilation has mostly desirable impact on the sub-surface chlorophyll. Since the OC assimilation updates chlorophyll only in the mixed layer, the impact on the sub-surface chlorophyll is the result of the model dynamical response to the assimilation. We demonstrate that the multi-platform assimilation combines the advantages of its components and always performs comparably to its best performing component.

than in the vertical, which means they are more dynamically stable in the horizontal than 123 in the vertical direction. For the gliders, it is of vital interest to understand the potentially 124 complex interaction between the physical and the biogeochemical data assimilation, or the 125 interplay between the different biogeochemical variables updated by the assimilative sys-126 tem. 127 In this study we extend the operational assimilative system on the NWE Shelf to 128 successfully produce a multi-platform reanalysis including both physical (satellite sea sur-129 face temperature, temperature and salinity from in situ platforms and an AlterEco glider) 130 and biogeochemical (total chlorophyll a and oxygen from an AlterECO glider, and chloro-131 phyll a from a satellite OC product) variables. The main focus of the paper is to assess 132 the impact of the different multi-platform assimilative system components (satellite vs  [2014]) with the biogeochemical model European Regional Seas Ecosystem Model (ERSEM, 156 Baretta et al. [1995]; Blackford [1997]; Butenschön et al. [2016]). We used measurements 157 from an AlterEco glider that operated in the central North Sea between May-August 2018 158 providing data for temperature, salinity, chlorophyll (derived from fluorescence) and oxy-159 gen concentrations. In multi-platform assimilation the glider data were complemented with  in Skákala et al. [2020]. The free run outputs have been analysed for the period of the 172 glider data availability (08/05-15/08, 2018). The assimilative runs used identical model 173 settings as the free run, only with the added assimilation components. The different as-similative runs compared in this study are (see also Table 1): a) physical data assimilation (satellite SST, temperature and salinity from EN4 data and the AlterEco glider), b) satel- 176 lite OC total chlorophyll a assimilation, c) AlterEco glider chlorophyll a assimilation, d) 177 AlterEco glider oxygen assimilation and e) multi-platform assimilation combining all the 178 data from a)-d). All the assimilative runs were started from the initial value conditions 179 produced by the free simulation for 08/05/2018.
where d ml is the mixed layer depth (MLD) and G(d) is the vertical grid spacing as a func- which assumes that for a suitable spatio-temporal binning the model and observational er- where the variances of the true state were estimated from the model outputs. This scheme  Table 2). For the two different chlorophyll observational products, the estimate of glider shown that the impact of grid-averaging on the biogeochemical reanalysis was negligible.

359
During each day the glider typically covered 3 model horizontal grid-cells and for each 360 model horizontal location the glider scanned nearly the full vertical water column.

361
The glider data (publicly available from www.bodc.ac.uk) were processed by the Na- The paper uses two metrics: a) model-to-observation bias (∆Q mo ) defined as where, as before, Q m are the model free run and Q o the observed concentrations (by the 392 observations we will automatically mean the glider data), and b) Bias-Corrected Root The BC RMSD metric is applied in two different contexts: as a "spatial BC RMSD" and 395 a "temporal BC RMSD".

396
In the case of spatial BC RMSD, we calculate for each day (t d ) the difference be-397 tween the model and the observed daily mean, which we call model-to-observation daily 398 bias: where Q m (t d ) and Q o (t d ) are the model free run and the observation data from the day t d , 400 and the model free run is taken only from the spatial locations visited by the glider (about as their daily biases: The spatial BC RMSD, ∆ S RD Q mo , is then obtained as a time-average of the daily BC RMSD, 405 i.e. averaging ∆ RD Q mo (t d ) through the glider data availability period (100 days for chloro-406 phyll and 53 days for oxygen): then applying equation 5 to those time-series, with bias understood as the model-to-observation 415 difference in the temporal mean of the time-series data: The temporal BC RMSD is designed to capture how the model represents the observed 417 phytoplankton phenology.

418
It should be noted that the metrics discussed in this section are used to measure 419 "the skill" of the assimilative runs by comparing the simulation outputs to the assimi-420 lated glider data, rather than to an independent validation data-set. There are two reasons 421 for this: firstly, to get sufficient validation data for the limited spatio-temporal region of 422 this study is nearly impossible, however, most importantly, this study has no ambition to 423 produce a skill-assessed reanalysis, its ambition is to test the impact of the assimilative

468
Since glider chlorophyll a data were assimilated across the whole water column, the 469 glider chlorophyll assimilation is also able to substantially improve the sub-surface chloro- reduced by almost 50% (Table 3 and Figure 6:D), the spatial BC RMSD by 60% (Table   474 3) and the temporal BC RMSD by 70% (Table 3). Unlike glider chlorophyll assimila-  Table 3) and spatial BC RMSD (by 15%, Table 3). Although the im-    to have relatively modest impact on chlorophyll bias, as well as spatial and temporal BC 560 RMSD (between 5-7%, Table 3). However, the impact of physical data assimilation on the 561 simulated phytoplankton could become more substantial within a strongly coupled system  Figure 9:A clearly shows that photo-585 synthesis is an important driver of the simulated oxygen, producing a large oxygen surge 586 in the mixed layer during the simulated late spring bloom. Some connection between oxy-587 gen and chlorophyll concentrations (a proxy for primary productivity) appears also in the 588 glider observations (Figure 9:B), with the peak in oxygen concentrations located in the 589 neighborhood of the glider deep chlorophyll maxima (Figure 3:B). As for chlorophyll, a 590 simple way to improve simulated oxygen is to assimilate the glider oxygen data into the 591 model ( Figure 10:D, Figure 11:H). Assimilating glider oxygen into the model reduces the 592 oxygen bias by 97%, temporal BC RMSD by 84% and spatial BC RMSD by 45% (Table   593 3). However, as in the case of chlorophyll, such assimilation has a limited spatial impact  Table 3).

629
The multi-platform chlorophyll re-analysis is dominated in the vicinity of the glider by the  will be able to cover only limited parts of the NWE Shelf and future multi-platform as-708 similative system will have to rely heavily on satellite data. . The panels display the skill of the following system components: physical data assimilation (grey color), satellite OC chlorophyll assimilation (orange), glider chlorophyll assimilation (light blue) and oxygen assimilation (brown). These components are compared with the multi-platform assimilative run (joint physical data, glider chlorophyll and oxygen, and satellite chlorophyll assimilation, green color), the free run (blue) and the glider observations (red).   e) The multi-platform assimilation (joint physical data, glider chlorophyll and oxy-727 gen, satellite OC chlorophyll assimilation) combines optimally the skill of its components 728 and always performs comparably to, or better than its best performing component.