The first source of uncertainty reflects the inherent bias of satellite-derived chlorophyll concentration and is assessed by comparing the S-NPP VIIRS chlorophyll to in situ fluorometric chlorophyll data. For the period from 2012 to 2014, the global chlorophyll from S-NPP VIIRS compared favorably to in situ chlorophyll (bias = 11.8%, semi-interquartile range = 27.9% and R = 0.86; Table 1).

The second source of uncertainty lies in how well the model simulates chlorophyll concentration. This source of uncertainty is assessed by comparing the chlorophyll concentration (Toggweiler et al., 1991) from the free-run model (no data assimilation but uses transient forcing conditions from MERRA) with the corresponding satellite ocean color data. For the period from 2012 until 2015, monthly chlorophyll concentration from the free-run model were significantly correlated to those from the satellite ocean color (S-NPP VIIRS, R = 0.72, p < 0.05; Table 1). The chlorophyll from the free-run model was on average within 27.87 ± 1.72% (average ± standard error) of the S-NPP VIIRS chlorophyll. Chlorophyll fields in the Equatorial Pacific showed agreement with satellite data (Figure 2). The model reproduces the main features observed by the satellite ocean color. The consistent positive bias in chlorophyll concentration in the Equatorial Pacific in the free-run model suggest that the upwelling in the Equatorial Pacific in the model is overestimated and therefore leads to higher chlorophyll concentration than those observed. The overprediction of the upwelling in the Equatorial Pacific in models has been suggested for some time (e.g., Toggweiler et al., 1991; Zheng et al., 2012). In some other areas, such as along the South America coastline as well as in the region of the Costa Rica Dome, the chlorophyll concentration from the free-run model was underestimated. This is most likely due to the nature of the reduced gravity circulation model. The model therefore does not include topographic effects, nor does it allow the representation of cross-shelf advection and convection.

In the Equatorial Pacific, the monthly chlorophyll concentration from a run assimilating S-NPP VIIRS chlorophyll data was significantly correlated (R = 0.95, P < 0.01; Table 1) and on average within 12.34 ± 0.52% of the S-NPP VIIRS chlorophyll concentration. The assimilation of satellite chlorophyll to provide the initial conditions used for the forecast is therefore an improvement over using the initial conditions provided by the free-run model without data assimilation. We therefore use this set-up to provide the initial conditions for the forecasting systems.

Finally the data used to force the model have their own inherent bias and uncertainties. While this is beyond the scope of this paper, we note that the bias in the forcing data used here have been assessed in other papers (e.g., Rienecker et al., 2011). By comparing the chlorophyll concentration from the free-run model using climatological MERRA forcing data compared to using transient MERRA data we can assess the improvements that such transient forcing data can provide to the system. The chlorophyll concentration from the free-run model using transient MERRA forcing data were considerably closer to the chlorophyll concentration from the S-NPP VIIRS (27.87 ± 1.72%) than the free-run model using climatological MERRA data (85.67 ± 2.77%, Figure 3). This indicates the advantage of using transient forcing data to further improve the initial conditions used for the forecasting system.

We assess the skill of our forecast by comparing each 9-month forecast to the observed chlorophyll concentration in the Equatorial Pacific from S-NPP VIIRS for the corresponding month. There was a consistent positive bias in the chlorophyll forecasted, as in the hindcast from the free-run model compared with S-NPP VIIRS (Figure 2). Of the 38 forecasts, the average percent difference between the forecasted chlorophyll and the S-NPP VIIRS chlorophyll varied between 23% (3 months lead time, the equivalent of 0.033 μg chl L⁻¹) and 30.7% (9 months lead time, the equivalent of 0.042 μg chl L⁻¹, Figures 4, 5). Except for the monthly chlorophyll concentration at 5 and 6-month lead time, the chlorophyll concentration from the forecasts were always significantly correlated to those from S-NPP VIIRS (data not shown). The highest correlation coefficient was observed at 8-month lead time (R = 0.53, p < 0.01).

To assess the uncertainties in our forecast, we utilize two deterministic skill metrics: ACC and RMSE. The ACC for the forecast was significant for the 1-month lead time (R = 0.33, P < 0.05) as well as for the 8- and 9-month lag forecast (R = 0.42 and R = 0.41 respectively, Table 2). This indicates that for these leads, the forecast chlorophyll had statistically the correct phasing when compared to those from S-NPP VIIRS. The spatial distribution of the anomaly correlation coefficient further reflects the overprediction of the upwelling in this Equatorial Pacific (Figure 6). While the forecasted chlorophyll concentrations at 1-month lead are significantly correlated with those from S-NPP VIIRS for the majority of the Equatorial Pacific, some areas in the upwelling tongue are not significant. The second skill metric, RMSE, increased from 0.040 μg chl L⁻¹ at 1-month lead to 0.047 μg chl L⁻¹ at 9-month lead forecast. These results suggest that while the phasing may have been reasonable at 8- and 9-month lag forecast, the amplitude of the signal was getting worse. Regardless, RMSE of 0.047 μg chl L⁻¹ is still very acceptable for a 9-month lag forecast. These results suggest some skill in forecasting the chlorophyll variability in the Equatorial Pacific especially at 1-month lag when the ACC is significant and the RMSE is at its lowest. For all forecasts, the chlorophyll concentrations were always within 30.7% of the chlorophyll concentration from S-NPP VIIRS. This is similar to the uncertainties reported for this instrument (semi-interquartile range of S-NPP VIIRS versus in situ chlorophyll = 27.9%).

In the Equatorial Pacific, the El Niño Southern Oscillation is the dominant source of interannual variability and has been shown to have a considerable impact of the biogeochemistry, including chlorophyll concentration and recruitment of higher trophic levels, in this region (e.g., Strutton and Chavez, 2000; Martinez et al., 2009). Forecasting El Niño events is the focus of many prediction centers. While the focus of assessments such as the North American Multi-Model Ensemble home has been on the skills in forecasting sea surface temperature, there has been very little work on forecasting biogeochemical variables such as chlorophyll using a dynamical system. The temporal evolution of the various forecasts in this study highlights the variability between the forecasts and our skills in predicting the decline in chlorophyll concentration that was observed in the Equatorial Pacific during the 2015 El Niño event (Figure 4). Starting in January 2015 the forecast suggested a decline in chlorophyll concentration that would reach a minimum in May 2015 (average of the 8 forecasts available for this month of 0.13 μg chl L⁻¹). The S-NPP VIIRS data observed this minimum 1 month later in June 2015 (0.13 μg chl L⁻¹). The chlorophyll concentration from S-NPP VIIRS then increased to reach a peak in August 2015 (0.14 μg chl L⁻¹). This increase in chlorophyll was also reflected in the various forecasts although it was overestimated. After August 2015, chlorophyll concentration declined reflecting the onset of the 2015 El Niño and the suppression of the upwelling in the Equatorial Pacific. This decline was also observed in the chlorophyll concentration from S-NPP VIIRS. Of the four forecasts available for September 2015, only one had predicted this decline. The other three forecasts predicted a decline but delayed by 1 month (chlorophyll started to decline in October 2015). For the four forecasts, September 2015 was their 6- to 9-month lead forecast which we previously showed had relatively low skills compared to the 1-month lead forecasts. In the last forecast (highlighted in red in Figure 3), September 2015 corresponded to its 6-month lead forecast and this forecast predicted particularly well the decline in chlorophyll concentration that occurred between August and December 2015 in the Equatorial Pacific in response to the El Niño event. The spatial distribution of the chlorophyll anomaly between December 2015 and March 2015 (first month of the last forecast available) coincides well with that from S-NPP VIIRS for the corresponding month (Figure 7). The area of negative anomaly in chlorophyll concentration along the South American coast is distinguishable in both the forecast and the S-NPP VIIRS chlorophyll data. The overestimation of the upwelling system in the forecast is also visible on this spatial representation of the chlorophyll anomalies. The temporal evolution of these various forecasts highlights the impacts that the atmospheric forcing data have on the forecast of chlorophyll. As the forecasts get closer to the El Niño event, the forecasted atmospheric and oceanographic physical forcing data have more skills and therefore lead to a better forecast in chlorophyll concentration. The forecast of chlorophyll in this region therefore relies heavily on the existence of accurate forecast of atmospheric forcing data. The initial conditions seem to play a more minor role in the forecasting skill for predicting chlorophyll in this region.

The uncertainties in the forecast of atmospheric and oceanic variables used to force the model play a critical role in our ability to provide a successful forecast. The skill of the variables produced by the GMAO forecasting system and that are used to force the model in forecast mode can also be a source of uncertainties and have been assessed in (Borovikov et al., in review). The SST anomaly correlation coefficient from the forecast in the tropical Pacific has a high correlation coefficient (R > 0.8) with the Reynolds SST for lead month 1–3 and remained above 0.6 by lag month 9 indicating significant (p < 0.05) skill. A case study of the El Niño event of 2015/2016 in (Borovikov et al., in review) suggested an overprediction of the magnitude in SST anomalies observed during the 2015/2016 El Niño event but was overall in good agreement with the conditions that were observed.

The forecast of chlorophyll concentration presented here is based on one single set of forecasting data while the forecasting system used at GMAO provides forecasts for several ensembles. Using ensemble forecasting instead of a single forecast might further improve our skill. Initial conditions can be perturbed in various ways to account for initial condition uncertainty. The uncertainty in the forecasted forcing data provided by GMAO could be accounted for by running with the various ensembles they provide for the variables used to force the biogeochemical forecast. Finally the model uncertainty could be accounted for using some stochastic parametrization at the sub-grid level such as the one used by the European Centre for Medium Range Weather Forecasts (Buizza et al., 1999).

Another source of uncertainty in our forecast is the assimilation methodology, the Conditional Relaxation Analysis Method used for bias correction for SST products (Reynolds, 1988) and applied here for chlorophyll (Gregg, 2008). This method does not utilize ensembles which can potentially improve the initial conditions for the forecast. It would also extend the memory of the assimilation, which appears to survive <2 months here and assist in the skill of the 1-month forecast. However, there is little evidence that the 2–9 month forecasts could benefit substantially from improved initial conditions, which are quite close to the S-NPP VIIRS chlorophyll as suggested in Table 1.

While these results suggest some skill in our ability to forecast chlorophyll concentration in the Equatorial Pacific, they also highlight potential weaknesses and avenues for improvements. The skill of the forecasting system relies as previously mentioned on the bias in the model's representation of physical and biogeochemical processes in the oceans, and the uncertainties in the forcing and assimilation data used. To further improve the forecasting system, each of these sources of bias and uncertainties needs to be assessed individually for weaknesses and possibilities for improvements. The range of applications of such a forecasting system, once properly set, can be extended for other variables. Applications include but are not limited to the prediction of Harmful Algal Blooms, fisheries, hypoxia/anoxia events, oil spills or the dispersal of pollutants. Prediction of temperature, ocean currents and velocities have for example been used for monitoring fisheries success, transport and spread of fish larvae, as well as seasonal fish migration (Johnson et al., 2005; Hobday and Hartmann, 2006; Bonhommeau et al., 2009). While the use of physical variables such as temperature, salinity and currents have been successfully used as covariates to explain distribution and catch rates of various species (e.g., Herron et al., 1989; Cole, 1999; Zagaglia et al., 2004; Bigelow and Maunder, 2007; Kaplan et al., 2016), these relationships can be limited since the behavior and recruitment of fish relies on changes in their prey concentration and composition. Accurate forecasts of the resources on which fish populations rely could provide the potential for strategic rather than reactive marine resource management during El Niño events for example. In the Equatorial Pacific, forecast of the effects of ENSO events on the physical conditions have been the subject of several studies starting in the 1980s (Cane et al., 1986). In the last two decades we have witnessed the development of two major El Niño events that had considerable impacts on both land and ocean conditions. The 1997-98 El Niño was particularly devastating for the ocean resources and led to the collapse of several fisheries and dramatic socio-economical repercussions for countries such as Peru. Anchovies, as well as other fisheries collapsed during both the 1982-83 and 1997-98 El Niño events. Forecasts such as the one presented here could therefore provide a framework to improve our management of resources during these events. Furthermore, the forecasting system presented here may provide a basis to expand the forecast from total chlorophyll to specific species including Harmful Algal Blooms. This could provide support for the management of many areas that need to monitor closely any development of harmful species in their waters. In the regions prone to Harmful Algal Blooms, such a forecast could also be used to improve the strategies to detect and manage most efficiently these events to minimize the repercussion on the human population and the associated economy.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.