Impact of the sea surface temperature forcing on hindcasts of Madden-Julian Oscillation events using the ECMWF model

This paper explores the sensitivity of hindcasts of the Madden Julian Oscillation (MJO) to the use of different sea surface temperture (SST) products as lower boundary conditions in the European Centre for Medium-range Weather Forecasts (ECMWF) atmospheric model. Three sets of monthly hindcast experiments are conducted, starting from initial conditions from the ERA interim reanalysis. First, as a reference, the atmosphere is forced by the SST used to produce ERA interim. In the second and third experiments, the SST is switched to the OSTIA (Operational Sea Surface Temperature and Sea-Ice Analysis) and the AVHRRonly (Advanced Very High Resolution Radiometer) reanalyses, respectively. Tests on the temporal resolution of the SST show that monthly fields are not optimal, while weekly and daily resolutions provide similar MJO scores. When using either OSTIA or AVHRR, the propagation of the MJO is degraded and the resulting scores are lower than in the reference experiment. Further experiments show that this loss of skill cannot be attributed to either the difference in mean state or temporal variability between the SST products. Additional diagnostics show that the phase relationship between either OSTIA or AVHRR SST and the MJO convection is distorted with respect to satellite observations and the ERA interim reanalysis. This distortion is expected to impact the MJO hindcasts, leading to a relative loss of forecast skill. A realistic representation of ocean–atmosphere interactions is thus needed for MJO hindcasts, but not all SST products – though accurate for other purposes – fulfill this requirement.


Introduction
The Madden-Julian Oscillation (MJO) is the major mode of intraseasonal variability (ISV) in the tropical atmosphere (Zhang, 2005). It is characterized by an Eastward propagation of regions of both enhanced and suppressed convection, mainly observed 20 over the Indian and the Pacific Oceans. The MJO is known to influence the Asian (e.g. Murakami, 1976;Yasunari, 1979) and Australian monsoon (Hendon and Liebmann, 1990), the evolution El Nino events (e.g. Kessler and McPhaden, 1995) and the weather regimes over the North Atlantic European region in winter (Cassou, 2008;Vitart et al., 2010). The simulation and the predicatibility of such intraseasonal and 25 seasonal weather regimes need an accurate representation of the MJO in the General 2536 accurate air-sea interactions through a good representation of the ISV and of the diurnal cycle of the SST. By focusing on the representation of the MJO in the Seoul National University atmospheric GCM, Kim et al. (2008) showed that high temporal SST frequency improved the simulation of the atmospheric ISV, of the propagation of the MJO and increased the MJO forecast skill. Kim et al. (2008) also argued that the 15 simulation of the MJO needs a good representation of the phase relationship between SST and convection as provided in coupled general circulation model.
In recent years the increase of number of satellite instruments has enhanced the developement of SST analysis products, such as those from the Group for High-Resolution Sea Surface Temperature (GHRSST, see Donlon et al., 2007, http://www. 20 ghrsst-pp.org/). Among these products is the recent 1/4 • daily OSTIA (Operational Sea Surface Temperature and Sea-Ice Analysis) SST reanalysis , which uses both satellite and in-situ data and spans the period January 1985-December 2007. The potential impact of such SST reanalysis in extended-range hindcast and atmosphereic reanalysis activities has to be assessed. For instance, the latest 25 ECMWF ERA interim atmospheric reanalysis (Dee et al., 2011) uses SST from different sources according to the considered period: the 1 × 1 • weekly NCEP 2d-var reanalysis from January 1981 to June 2001 (Reynolds et al., 2002), the 1 × 1 • weekly NCEP OIv2 SST reanalysis from July 2001 to December 2001 (Reynolds et al., 2002), the daily 1/2 • Printer-friendly Version

Interactive Discussion
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Real Time Global (RTG) SST analysis from January 2002 to January 2009 (Gemmill et al., 2007) and the 1/20 • daily OSTIA from February 2009 onwards (Donlon et al., 2011). Before 1981 and the satellite era, the ECMWF reanalyses use the HADISST1 dataset of monthly SST values produced by the Met Office (Rayner et al., 2003). This work is an attempt to assess the impact of different kinds of SST forcings on the 5 prediction of past MJO events in atmosphere-only mode using a recent cycle of the IFS. The winter 1992/1993 MJO is used as benchmark case at ECMWF as in Woolnough et al. (2007). First the impact of the temporal resolution of SST on the MJO prediction is investigated. The skill of the IFS in predicting the 1992-1993 MJO is assessed for different SST fields: persisted SST anomalies, weekly observed SST used in ERA 10 interim, monthly ERA interim SST (to simulate the pre-sattelite era), daily OSTIA SST reanalysis. Then this study particularly insists on the experiments using the current ECMWF observed SST and the OSTIA reanalysis to highlight the sensitivity of the intra-seasonal prediction to "realistic" SST forcing. In the following, Sect. 2 describes the experiment settings, the SST products and the resulting MJO scores. Section 3 15 focuses on the difference between the forecast using OSTIA SST and observed SST from ERA interim. Section 4 discusses the outcomes of the experiments and Sect. 5 draws the conclusions of this study.  (Dee et al., 2011). A skin layer scheme has been implemented in the IFS to simulate the diurnal variations of SST (see Zeng and Beljaars, 2005;Takaya et al., 2010). Four MJO experiments (see Table 1) are conducted in atmosphere-only mode where the atmosphere is forced by: (i) persisted SST anomalies (perSSTa), (ii) SST from ERA interim (ERAi), (iii) monthly SST (MTH) and (iv) OSTIA 5 SST. These SSTs corresponds to the foundation SST defined by the GHRSST as the temperature at the base of the diurnal thermocline.

SST products
Persisted SST anomalies are obtained by adding ERA interim SST anomalies (against the climatology) of the starting date to the SST climatology corresponding to the fore-  , 2011), which is independent of ERA interim. This analysis combines the information from in situ data (from ships and buoys), from the AVHRR satellite and 20 from the Along Track Scanning Radiometer (ATSR) instruments on board of the ERS-1, ERS-2 and ENVISAT satellites.

Diagnostic procedure
The skill of the monthly forecasting system in predicting the MJO is evaluated according to the method described in Wheeler and Hendon (2004 (EOF) analysis of the anomalies (with respect to the 1991-2003 climate) of the zonal wind at 200 hPa and 850 hPa and of the outgoing longwave radiation (OLR) averaged between 10 • S and 10 • N. Wheeler and Hendon (2004) showed that most of the MJO variability is described by the two first components of the combined EOF analysis. Their longitudinal patterns can represent all the active and suppressed phases of the MJO 5 ( Fig. 1) over its eastward propagation. Negative OLR extrema reflect the position of the convective centre of the MJO. According to the sign of the associated Principal Component (PC), the convective centre on EOF1 is located over the Maritime Continent (PC1 > 0) or over the Western Hemisphere and Africa (PC1 < 0). On EOF2, the convection is over the Pacific Ocean (PC2 > 0) or over the Indian Ocean (PC2 < 0).

10
The recommended score of the MJO forecast relies on the correlation of the monthly ensemble-mean forecasts with the two first PCs of the combined EOFs estimated from the ERA interim atmospheric reanalysis. According to Woolnough et al. (2007), two MJO events occur between mid-December 1992 and February 1993. The 47 starting dates of the experiments include all the phases of these MJO events as identified by 15 the combined EOF analysis. Plus, each forecast captures each phase of the MJO at least once.

Results
The correlations of the ensemble-mean forecast with the two principal components of 20 the combined EOF are estimated for all the experiments (Fig. 2). The forecast skill is considered as significant for correlations higher than 0. 18 implying a gain of at least 4 days of MJO forecast skill on both EOFs. Such gain was expected considering that ERAi and OSTIA SST -being real-time observed fields -have an ISV signal that is closer to the truth than the climatological signal of the persisted SST anomalies. Another expected result is the degradation of the forecast when using monthly SST. The MTH experiment scores are sometimes even worse 5 than perSSTa and, when comparing to ERAi and OSTIA, they show a loss of significant forecast skill of at least 3 to 5 days on EOF 1 and 2, respectively. Such performance has to be expected when reforecasting similar MJO events before 1981 (pre-satellite era). ERAi and OSTIA are the two best experiments. Their scores are similar until day 16 on EOF1 and 18 on EOF2, then ERAi provides a gain of one day of significant skill 10 on EOF1 and at least more than 3 on EOF2. Forcing with weekly ERAi SSTs is better at predicting the MJO of the winter 1992-1993 than forcing with daily OSTIA SSTs. This result is unexpected because, according to Kim et al. (2008) and Klingaman et al. (2008), the daily frenquency should have a positive impact on the representation of the tropical ISV and thus on the forecast of the MJO.

Outcomes
As providing higher ISV to the atmosphere seems to improve significantly the MJO prediction in the case of the ERAi and OSTIA experiments, one can think of an alternative to forcing by persisted SST anomalies or monthly SSTs. According to works from Woolnough et al. (2007) and Vitart et al. (2007) coupling the atmosphere to the ocean 20 would be a good candidate in the case of the extended range forecast or from the reanalysis point of view. Addressing these aspects is nevertheless beyond the scope of our study. More interesting is the fact that forcing with weekly ERAi SSTs provides better MJO prediction than forcing with daily OSTIA SSTs. To visualise how the two forced experi- Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | averaged for starting dates when the convective centre of the MJO is over the Indian Ocean. In the analysis, the convective centre of the MJO (negative OLR anomalies) propagates from the Indian to the Central Pacific Ocean and is followed by a phase of suppressed convection (positive OLR anomalies) a few days later. The ERAi experiment simulates correctly this propagation but the MJO active and suppressed phases 5 are much weaker than in the analysis. The weakening is particarly marked when the convection reaches the Maritime Contient that is known as a barrier for the MJO simulation (Inness et al., 2003). In the OSTIA experiment, the MJO signal is even weaker over the Maritime Continent and its eastern propagation is hardly visible. The difference between the ERAi and OSTIA experiment only comes from the SSTs forcing 10 the atmosphere. The relative impact of these two SST fields on the MJO prediction is investigated in the next sections.

OSTIA versus ERAi SSTs: mean state and daily variability
Apart from the horizontal resolution of the source fields (before the interpolation on 15 the atmospheric grid), the main differences between OSTIA and ERAi SSTs come from their mean state and the additional noise associated to the daily frequency of the OSTIA product. On average over the winter 1992-1993, the OSTIA SSTs are overall colder than ERAi SSTs by 0.18 • C in the Tropics. Apart from some warmer patches, OSTIA SSTs are particularly colder (sometimes by more than 0.4 • C) in the western 20 part of the Maritime Continent, in the Pacific cold tongue and in the Tropical Atlantic ( Fig. 4) Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | agreement is not observed at every TAO station but supports the accuracy of the OS-TIA product.

OSTIA versus ERAi SSTs: phase relationship between SST and convection
This section investigates the difference in the phase relationship between SST and atmospheric convection. This phase realtionship is estimated in the Indian Ocean over 5 the winters (December-February) 1985-2006: from ERAi and OSTIA SSTs fields and OLR (indicative of the convection) from the ERA interim reanalysis and from National Oceanic and Atmospheric Administration (NOAA) daily interpolated OLR (see Liebmann and Smith, 1996). The NOAA interpolated OLR is produced from the NOAA satellite retrievals on a 2.5 • × 2.5 • grid and is available from 1979 onwards. The phase 10 relationship between SST and convection is produced from filtered SST and OLR anomalies averaged in the Indian Ocean box 5 to 19 days for PC1 and 16 days for PC2, which is substantially less than the 20 days and more obtained for the only 1992-1993 case. The difference between the ERAi and OSTIA experiments is weaker than for the 1992-1993 case with rarely more than 1 day of loss of forecast skill when using OSTIA SST instead of ERAi SST. The scores nevertheless confirm that forcing the atmopshere by OSTIA SST is less efficient than 5 forcing by ERAi SST (Fig. 7) in predicting the MJO.

Phase relationship between SST and convection in the experiments
The 22-winter experiments provide enough data to investigate the phase relationship between the OSTIA and ERAi SST and the atmospheric convection (OLR) according to the forecast lead time. This relationship is estimated in a similar way as in Sect. 4.2.

10
The interannual variability in each 32-day forecast is removed by substracting its 32-day mean. The intraseasonal variability is then extracted by applying a 5-day running mean in each forecast segment. The evolution of the phase relationship between the SST and OLR in the Indian Ocean (5 • S-5 • N, 60 • -95 • E) according to the lead time in the two forced experiments is compared to the equivalent in the ERA interim analysis. As seen 15 in Sect. 4.2 (Fig. 5), the analysis shows a near-quadrature phase relationship (Fig. 8).
Extremum correlations occur around 7-10 days according to the considered forecast week. Both the forecast forced by OSTIA SST and ERAi SST produce similar phase relationships that are, though sometimes weakened, overall close to the analysis until week 3 of the forecast. The forecast forced by OSTIA is nevertheless slightly shifted 20 toward negative lags in week 1 of the forecast. The phase relationship is recovered in weeks 2 and 3 but with lower correlations than in the ERAi experiment when the lag is negative.

Discussion and conclusion
This study suggests that the relative loss of MJO forecast skill when forcing with OSTIA SST is related to an inaccury in the phase relationship between OSTIA SST and the atmospheric convection produced in the model (Fig. 8). A similar inaccuracy has also been reported between OSTIA SST and the convection derived from satellite observa-5 tions (Fig. 5). From the initial state of the MJO prediction using OSTIA SST, the atmosphere and its lower boundary conditions are not as consistent as they are when using ERAi SST. This is most likely linked to the fact that the initial state of the atmospheric model comes from the ERA interim reanalysis that has been produced using the ERAi SSTs as boundary conditions for the atmospheric model. One can easily expect an ini-10 tialization shock that would deteriorate the prediction. Moreover, the OSTIA SSTs here come from a 1/4 • resolution product that is interpolated on the T159 atmospheric grid for the purpose of our experiments. Even smoothed by the interpolation, the resulting fields are spatially much noisier than ERAi SSTs. Noisy features may generate air-sea interactions weakening the MJO signal in a low resolution atmosphere starting from an 15 initial state produced by using smooth ERAi SST fields. The Maritime Continent being a barrier to the MJO prediction, an initially weakened MJO signal will have difficulties to propagate over and past this barrier as described on Fig. 3 in the OSTIA experiment. A way to assess the potential importance of the initial bias of the atmosphere toward ERA interim, would be to produce a similar atmospheric reanalysis using OSTIA SSTs daily OSTIA SSTs show that SSTs with accurate intraseasonal variability significantly improved the prediction. The MJO prediction shows substantial sensitivity to these two SST forcings. The prediction forced by OSTIA SST shows a loss of forecast skill of several days when compared to the prediction forced by ERAi SST. Sensitivity experiments show that this deterioration is linked neither to the difference in mean state between 5 OSTIA and ERAi SSTs nor to the OSTIA daily variability. Additional diagnostics show that, the OSTIA product provides an anoumalously weak and shifted phase relationship on the intraseasonal time-scale between SST and convection from the beginning of the prediction. This phase relationship is never totally recovered, leading to the simulation of a weakened MJO signal with increased difficulties to propagate. Though close 10 to in-situ SST observations, the consistency of the OSTIA product with the ECMWF atmospheric model is not optimal.  , 132, 1917-1932, 2004.  Soc. Jpn, 57, 225-242, 1979. 2536 Zeng, X. and Beljaars, A.: A prognostic scheme of sea surface skin temperature for modeling and data assimilation, Geophys.