**3. Results**

#### *3.1. How Does the Operational CMEMS IBI Forecast Model Reproduce Salinity Fields?*

In this section, and before analyzing specific results from the IBI model sensitivity tests performed, it is shown how the CMEMS IBI operational model system reproduced salinity fields for the study year (2018). NARVAL toolbox provided a wealth of outcomes and quality indicators on a monthly basis that allowed to quantify CMEMS-IBI prognostic capabilities to reproduce the salinity field during such year. Equally, "parent-son" model intercomparisons were regularly conducted with the purpose of checking the consistency of the CMEMS global solution and the regional downscaled IBI model solution, identifying any potential problem that might be inherited from the coarser global system. In this context, Lorente et al. [17] previously showcased the discrepancies in coastal areas during 2018 between the CMEMS IBI and its parent solution (CMEMS GLOBAL) due to the differences in both the horizontal resolution and the freshwater forcing implemented in their respective operational chains.

The CMEMS operational IBI performance appeared to be rather consistent, especially in open deep waters (Figure 4). The qualitative resemblance of yearly averaged maps of SSS for 2018 between the IBI model product and SMOS estimations was significantly high, with slight differences arising in nearshore shallower areas, especially in the African coastal upwelling region (Figure 4a,b).

With the aim of assessing IBI accuracy in the entire water column, quality-controlled salinity profiles collected by Argo floats were regularly used as a benchmark (Figure 4c–f). Monthly maps of skill metrics were computed by NARVAL toolbox for several depth layers and also for the entire full profiles, as exposed in Figure 4c,d. As it can be observed, moderate RMSE and high correlation values were obtained for March 2018, with most of the values emerging in the range [0–0.5] PSU and [0.71–0.93], respectively. Complementarily, monthly qualitative comparisons were conducted with a focus on specific subregions (Figure 4e,f). Here, the salinity profiles observed and modeled in the Gulf of Biscay are shown (the BISCA subregion displayed in Figure 2). Model outputs were vertically interpolated into the observation depth levels to facilitate the visual comparison. Likewise, averaged salinity profiles (solid black lines) were calculated to infer the main features of the water column in this area. According to the values for specific levels (in blue), the model-observations accordance was relevant for March 2018. Reduced differences were detected in the upper layer, where higher spatial-temporal variability of the salinity field is observed, lying between 33 and 36 PSU. Monthly statistical results for the rest of the year (not shown) illustrated that CMEMS IBI performance in off-shelf waters was consistent, operating within acceptable ranges all the year-long.

Since accuracy of SMOS remote-sensed estimations is higher in open offshore waters (these products showing a more limited precision in coastal areas mostly due to the lack of valid observations—this is so especially in the operational products and somehow minimized in the reprocessed ones) and Argo observations were constrained to offshore waters beyond the continental shelf, ancillary validation exercises with coastal buoys were required to objectively evaluate CMEMS IBI performance in the periphery of ROFI areas, where impulsive-type freshwater discharges might be noticeable. Using surface salinity observations from the coastal moorings available along the northwestern Iberian shelf (M5–M12 mooring buoys, depicted in Figure 2), a contingency table is served, not only as a summary of the most relevant SSS drop events occurred during 2018, but also as a qualitative overview of the CMEMS IBI model ability to adequately reproduce them (Figure 5a).

**Figure 4.** 2018 yearly averaged maps of sea surface salinity (SSS) simulated by CMEMS IBI (**a**) and derived from remotely sensed SMOS data (**b**). Monthly maps (March 2018) of skill metrics derived from the comparison of entire salinity profiles observed by Argo floats and modelled by CMEMS IBI: root mean squared error (**c**) and correlation (**d**). Daily salinity profiles collected by Argo floats (**e**) and IBI (**f**) in the Gulf of Biscay (BISCA zone) during March 2018. Averaged profiles exposed in solid black lines. Mean depth values gathered in blue.

**Figure 5.** (**a**) Contingency table focused on significant sea surface salinity (SSS) drops detected during 2018 at coastal buoys moored around the Iberian Peninsula (buoys M2–M12 used, see Figure 2 for location). Hit, miss and false alarm events are highlighted in green, purple and yellow colors, respectively. (**b**) Six-month observed (blue dots) and IBI modeled (red line) SSS timeseries at Silleiro buoy (M12). Relevant SSS drops marked with colored dashed squares. (**c**) Three-month timeseries comparison of SSS observed at Cabo de Peñas buoy (M7) and modelled by CMEMS-IBI at the closest grid point.

As it can be deduced from the contingency table, CMEMS IBI was able to represent 69% (24 of 35) of the episodes categorized here, with variable degreed of precision. By contrast, almost 50% (22 of 46) of the forecasted SSS drop events constituted false alarms, which were not confirmed by the in-situ observation. To better illustrate this validation approach, two examples at 2 buoys' locations are provided (Figure 5b,c). In the first case, the 6-month SSS timeseries at Silleiro buoy (M12) location revealed a series of events, categorized as modelled hit and miss event (Figure 5b), chronologically described hereinafter:


In the second case (Figure 5c), the 3-month comparison of SSS at Cabo de Peñas buoy (M7) location showed the following timeline features (hits and a miss event):


#### *3.2. IBI Model Sensitivity to Changes in Freshwater Coastal and River Forcing*

This section shows the main results of the IBI modeling scenarios performed to evaluate the regional impact on salinity after using different river/coastal freshwater

forcing. In order to assess the four different IBI scenarios performed and to identify the main sources of uncertainty linked to the use of different river runoff forcing, the modelled outcomes have been validated with different in-situ salinity observations.

Figure 6 shows the seasonal SSS of IBI\_REF, and the mean difference with the other simulations. In winter, that is the first 3 months after the start of the runs, the differences in the surface salinity field between the various simulations are limited to the shelf, and more specifically to the ROFI areas. Due to the higher river discharge of LAMBDA, IBI\_LAM and IBI\_NOR are fresher than IBI\_REF. The same patterns are found in autumn. The small differences observed offshore at this time come from the expected propagation of the differences within the domain after 12 months of simulation but confirm that there is no noticeable bias appearing with any of the forcing sets. In spring and summer, the river discharge of LAMBDA is stronger than the other sources (as seen in Figure 3) and the fresher water masses of the two simulations forced by LAMBDA (IBI\_LAM and IBI\_NOR) extend to the open ocean, especially in the BISCA area. The Vilaine and Loire river system (at 47◦ N) impacts the surface and subsurface layers all year-long. The southernmost section of BISCA is the most sensible to the impact of the coastal runoff, as shown by the higher salinity of IBI\_NOR in this area.

The simulation forced by a climatology (IBI\_CLM) is saltier than the other simulations (+0.05 PSU over the IBBIS domain compared to IBI\_REF) in winter. This emphasizes the fact that using a climatology vs. realistic daily river discharge significatively changes the salinity budget on the shelf.

These mean SSS seasonal snapshots from the test simulations show differences in the salinity patterns. Furthermore, it is needed to validate the simulated salinity fields against observations in order to assess the impact of these changes in the freshwater forcing.

The comparison with observed salinity timeseries from moorings shows that all the simulations globally manage to accurately reproduce the spatial-temporal variability of salinity at the coast and at the shelf-break, within the IBBIS domain (Figure 7). The correlations between mooring measurements and model scenarios outputs are most of the time above 0.6, and the standard deviation below 0.5 PSU. The capacity of the IBI model to reproduce the salinity is more explained by the geography and resolution of the local dynamics than by the parametrization of the river forcing: it can indeed be seen on the Taylor diagram (Figure 7a) that points are gathered by station, rather than being gathered by simulations. The simulation forced 6 months by the climatology differs most of the time from the others: it may be explained by the impact of a diurnal vs. climatological freshwater discharge forcing or by the shorter length of the simulation (6 months vs. 1 year).

As each subregion features its proper dynamics and local processes, the assessment of the salinity variability and fronts location, and the assessment of the impact of the river discharge forcing on the IBI configuration, have been conducted separately in BISCA, WISHE and CADIZ areas.
