Mediterranean Production Centre MEDSEA_ANALYSIS_FORECAST_WAV_006_011

Transcription

1 Mediterranean Production Centre Contributors: Anna Zacharioudaki, Michalis Ravdas, Gerasimos Korres, Emanuela Clementi (Validator) Approval date by the CMEMS product quality coordination team: N/A 1

2 Page 2/ 40

3 CHANGE RECORD When the quality of the products changes, the QuID is updated and a row is added to this table. The third column specifies which sections or sub-sections have been updated. The fourth column should mention the version of the product to which the change applies. Issue Date Description of Change Author Validated By 1.0 February 2017 All Creation of the document Anna Zacharioudaki Emanuela Clementi (PQ Responsible) September 2017 II.1 II.2 II.3 V VI Change of the horizontal Anna Zacharioudaki resolution of the surface currents forcing in the Mediterranean Sea (from 1/16º to 1/24º) Emanuela Clementi (PQ Responsible) Page 3/ 40

4 Table of contents Executive summary... 5 Products covered by this document... 5 Summary of the results... 5 Estimated Accuracy Numbers... 6 Production system description... 8 Production centre details... 8 Description of the Med-waves WAM modelling system...10 Upstream data and boundary condition of the WAM model...13 Validation framework Validation results Significant wave height Mean Wave Period System s Noticeable events, outages or changes Quality changes since previous version References Page 4/ 40

5 EXECUTIVE SUMMARY Products covered by this document This document describes the quality of the analysis and forecast nominal product of the wave component of the Mediterranean Sea:. The product includes: 2D hourly instantaneous fields of: spectral significant wave height (Hm0), spectral moments (1,0) wave period (Tm-10), spectral moments (0,2) wave period (Tm02), wave period at spectral peak / peak period (Tp), mean wave direction from (Mdir), wave principal direction at spectral peak, stokes drift U, stokes drift V, spectral significant wind wave height, spectral moments (0,1) wind wave period, mean wind wave direction from, spectral significant primary swell wave height, spectral moments (0,1) primary swell wave period, mean primary swell wave direction from, spectral significant secondary swell wave height, spectral moments (0,1) secondary swell wave period, mean secondary swell wave direction from. Output data are produced at 1/24º horizontal resolution. Summary of the results The quality of the V3 MED-MFC-waves system of analysis and forecast is assessed over a 1 year period (2014) by comparison with in-situ and satellite observations. The main results of the product quality assessment are summarized below: Spectral Significant Wave Height (Hm0): Overall, the significant wave height is accurately simulated by the model. Considering the Mediterranean Sea as a whole, the typical difference with observations (RMSD) is 0.21 m with a bias ranging from m (3.7%) to m (5.5%) depending on the source of observations the model is compared against. In general, the model somewhat underestimates the observations for wave heights below 4 m whilst it mostly converges to the observations for higher waves. Its performance is better in winter when the wave conditions are well-defined. Spatially, the model performs optimally at offshore wave buoy locations and well-exposed Mediterranean subregions. Within enclosed basins and near the coast, unresolved topography by the wind and wave models and fetch limitations cause the wave model performance to deteriorate. Spectral moments (0,2) wave period (Tm02): The mean wave period is reasonably well simulated by the model. The typical difference with observations (RMSD) is 0.7 s and is mainly caused by model bias which has a value of s (12%). In general, the model underestimates the observed mean wave period and exhibits greater variability than the observations. A relatively larger model underestimate is found for mean wave periods below 4.5 s. Model performance is a little better in winter when wave conditions are well-defined. Spatially, the model somewhat overestimates the highest mean wave period values in the western Mediterranean Sea, west and south of France. Otherwise, model underestimate is widespread. Similarly to the wave height, the model performance is best at wellexposed offshore locations and deteriorates near the shore mainly due to fetch limitations. Page 5/ 40

6 Other variables: No observations are available for all other variables except for the wave period at spectral peak / peak period (Tp) and the mean wave direction from (Mdir). Qualification of these two variables will be considered in the near future. Generally, wave height variables are expected to be of similar quality to Hm0 and wave period variables to Tm02. Stokes drift quality is expected to be a function of both Hm0 and Tm02. Estimated Accuracy Numbers Estimated Accuracy Numbers (EANs), that are the mean and the RMS of the differences (RMSD) between the model and in-situ or satellite reference observations, are provided in Table 1 and Table 2 below. In following sections, "Mean" and "RMS" are referred to as "BIAS" and "RMSD" in alignment with metrics' name conventions within CMEMS. EANs are computed for: Significant Wave Height (SWH): refers to the "spectral significant wave height (Hm0)" Mean Wave Period (MWP): refers to the "spectral moments (0,2) wave period (Tm02)" The observations used are: in-situ observations from moored wave buoys obtained from the CMEMS INSITU_MED_NRT_OBSERV dataset, available through the CMEMS In Situ Thematic Assemble Centre (INS-TAC) satellite altimeter observations from a merged altimeter wave height database setup at CERSAT - IFREMER Figure 1: Mediterranean Sea sub-regions for qualification metrics The EANs computed for the V3 version of the CMEMS Mediterranean Sea wave modelling system are based on the simulation of the system in hindcast mode for a period of 1 year from January 2014 to December They are computed for the Mediterranean Sea as a whole and for 17 sub-regions from which 1 is in the Atlantic Ocean and 16 in the Mediterranean Sea (Figure 1): (atl) Atlantic, (alb) Alboran Sea, (swm1) West South-West Med, (swm2) East South-West Med, (nwm) North West Med, (tyr1) Page 6/ 40

7 North Tyrrhenian Sea, (tyr2) South Tyrrhenian Sea, (adr1) North Adriatic Sea, (adr2) South Adriatic Sea, (ion1) South-West Ionian Sea, (ion2) South-East Ionian Sea, (ion3) North Ionian 3, (aeg) Aegean Sea, (lev1) West Levantine, (lev2) North-Central Levantine, (lev3) South-Central Levantine, (lev4) East Levantine. SWH and MWP vs in-situ observations: full MED SWH MWP Mean RMS Table 1: EANs of SWH and MWP evaluated for the V3 system for a period of 1 year (2014) for the full Mediterranean Sea: SWH-3H-CLASS2-MOOR-BIAS-MED, SWH-3H-CLASS2-MOOR-RMSD-MED, MWP3H-CLASS2-MOOR-BIAS-MED, MWP-3H-CLASS2-MOOR-RMSD-MED in Table 4, Section III. Units are in metres for SWH and in seconds for MWP. SWH vs satellite observations: full MED and sub-regions MED atl alb swm1 swm2 nwm tyr1 tyr2 ion1 ion2 ion3 adr1 adr2 lev1 lev2 lev3 lev4 aeg Mean RMS Table 2: EANs of SWH evaluated for the V3 system for a period of 1 year (2014) for the full Mediterranean Sea and the different sub-regions shown in Figure 1: SWH-CLASS4-ALT-BIAS-MED, SWHCLASS4-ALT-BIAS-MED, SWH-CLASS4-ALT-BIAS-<REGIONS>, SWH-CLASS4-ALT-RMSD-<REGIONS> in Table 4, Section III. Units are in metres. Page 7/ 40

8 PRODUCTION SYSTEM DESCRIPTION Production centre details PU: HCMR, Greece Production chain: Med-waves External product (2D): spectral significant wave height (Hm0), spectral moments (-1,0) wave period (Tm-10), spectral moments (0,2) wave period (Tm02), wave period at spectral peak / peak period (Tp), mean wave direction from (Mdir), wave principal direction at spectral peak, stokes drift U, stokes drift V, spectral significant wind wave height, spectral moments (0,1) wind wave period, mean wind wave direction from, spectral significant primary swell wave height, spectral moments (0,1) primary swell wave period, mean primary swell wave direction from, spectral significant secondary swell wave height, spectral moments (0,1) secondary swell wave period, mean secondary swell wave direction from. Frequency of model output: hourly (instantaneous) Geographical coverage: ºW ºE ; ºN ºN Horizontal resolution: 1/24º Vertical coverage: Surface Length of forecast: 5 days Frequency of forecast release: Daily Analyses: No Hindcast: Yes Frequency of analysis release: N/A Frequency of hindcast release: Daily The wave forecasts for the Mediterranean Sea are produced by the HCMR Production Unit by means of the WAM wave model (described below). The Med-waves system is run once per day starting at 12:00:00 UTC. It produces 5-day (120 h) forecast fields initialized by a 1-day (24 h) hindcast. A schematic of the Med-waves operational cycle is shown in. The Med-waves system integration is composed of several steps: 1. Upstream Data Acquisition, Pre-Processing and Control of: ECMWF (European Centre for Medium-Range Weather Forecasts) NWP (Numerical Weather Prediction) atmospheric forcing and CMEMS Med MFC V3 and Global MFC V2 currents 2. Hindcast/Forecast: WAM produces one day of hindcast and 5 days of forecast. 3. Post processing: the model output is processed in order to obtain the products for the CMEMS catalogue Page 8/ 40

9 4. Output delivery Figure 2: Schematic of the Med-waves operational cycle: 120-h forecasts are initialized daily with initial conditions (IC) taken from the hindcast simulation of the previous day. Page 9/ 40

10 Description of the Med-waves WAM modelling system The waves component of the Mediterranean Forecasting Centre (Med-waves) is providing short-term wave forecast for the Mediterranean Sea at 1/24º horizontal resolution since year It consists of a wave model grid implemented over the whole Mediterranean Sea at 1/24º horizontal resolution nested within a coarser resolution wave model grid implemented over the Atlantic Ocean. Med-waves is based on the state-of-the-art third-generation wave model WAM Cycle (Günther and Behrens, 2012) which is a modernized and improved version of the well-known and extensively used WAM Cycle 4 wave model (WAMDI Group, 1988; Komen et al., 1994). Cycle has been released during MyWave ( A pan - European concerted and integrated approach to operational wave modelling and forecasting a complement to GMES MyOcean services ) EU FP7 Research Project and is freely available to the entire research and forecasting community. WAM solves the wave transport equation explicitly without any presumption on the shape of the wave spectrum. Its source terms include the wind input whitecapping dissipation, nonlinear transfer and bottom friction. The wind input term is adopted from Snyder et al. (1981). The whitecapping dissipation term is based on Hasselmann (1974) whitecapping theory. The wind input and whitecapping dissipation source terms of the present cycle of the wave model are a further development based on Janssen s quasi-linear theory of wind-wave generation (Janssen, 1989; Janssen, 1991). The nonlinear transfer term is a parameterization of the exact nonlinear interactions as proposed by Hasselmann and Hasselmann (1985) and Hasselmann et al., (1985). Lastly, the bottom friction term is based on the empirical JONSWAP model of Hasselmann et al. (1973). The Med-waves set-up includes a coarse grid domain with a resolution of 1/6º covering the North Atlantic Ocean from 75ºW to 10ºE and from 10ºN to 70ºN and a nested fine grid domain with a resolution of 1/24º covering the Mediterranean Sea from ºW to ºE and from ºN to ºN. The areas covered by the two grids are shown in. The bathymetric map has been constructed using the GEBCO bathymetric data set for the Mediterranean Sea model and the ETOPO 2 data set (U.S. Department of Commerce, National Oceanic and Atmospheric Administration, National Geophysical Data Centre, minute Gridded Global Relief Data) for the North Atlantic model. In both cases mapping on the model grid was done using bilinear interpolation accompanied by some degree of isotropic laplacian smoothing. The Mediterranean Sea model receives from the North Atlantic model full wave spectrum at hourly intervals at its Atlantic Ocean open boundary. The latter model is considered to have all of its four boundaries closed assuming no wave energy propagation from the adjacent seas. This assumption is readily justified for the north and west boundaries of the North Atlantic model considering the adjacent topography which restricts the development and propagation of swell into the model domain. The choice of the south boundary location is less obvious and is based on a number of studies which agree that no important swell energy is expected to propagate northwards from geographical areas south of 10ºN. Specifically, according to Semedo et al. (2011), a swell front present in all seasons can be identified in the Atlantic Ocean within the latitude band from 15ºS (Dec-Jan-Feb) to 15ºN (Jun-JulAug). Young (1999) suggests this swell front never migrates north of the equator. The relatively narrow geometry of the Atlantic restricts propagation of Southern Ocean swell into the Northern Hemisphere. According to Alves (2006) storms within the extratropical South Atlantic ocean (below 40ºS) typically propagate to the east spreading swell energy to the Indian Ocean. As for the Atlantic tropical areas, storms rarely evolve in the south band (between 20ºS and the equator) while in the north tropical Page 10/ 40

11 band (between the equator and 20ºN) summer storms move mostly westwards. During winter, the north tropical band can be affected by eastward propagating North Atlantic extratropical storms generating swells that propagate to the southeast (Alves, 2006). The wave spectrum is discretized using 32 frequencies, which cover a logarithmically scaled frequency band from Hz to Hz (covering wave periods ranging from approximately 1s to 24s) at intervals of = 0.1, and 24 equally spaced directions (15 degrees bin). ECMWF FC/AN Surface winds 1/8 x 1/8 MED/GLO MFCs Surface currents 1/24 x 1/24 (MED) 1/12 x 1/12 (GLO) The Mediterranean model runs in shallow water mode considering wave refraction due to depth and currents in addition to depth induced wave breaking. The North Atlantic model runs in deep water mode with wave refraction due to currents only. The North Atlantic model additionally considers wave energy damping due to the presence of sea ice. A model grid point is considered to be a sea ice point if the ice fraction at that point exceeds 60%. At all sea ice points the wave energy is set to zero. WAVE MODEL WAM (24 dir & 32 freq bins) Hourly output of wave parameters 1/24 x 1/24 1/6 x 1/6 + cg Ν ( x, t, f, θ ) = Sin + S dis + S NL + Sbot x t Figure 3: Bathymetry and model configuration of Med-waves Following ECMWF (ECMWF, 2015), the tunable whitecapping dissipation coefficients been altered from their default values. Specifically, the values of ( and have default) and ( = 0.6 default) have been adopted. The aim of this tuning is to produce results which are in good agreement with data on fetch-limited growth and with data on the dependence of the surface stress on wave age. The system is forced by 10m above sea surface wind fields obtained from the ECMWF Integrated Forecasting System (IFS) at 1/8º resolution. The temporal resolution of the wind forcing is 3-h for the first 3 days of the forecast and 6-h for the rest of the forecast cycle. The wind is bi-linearly interpolated onto the model grids. Sea ice coverage fields are also obtained from ECMWF at the same horizontal resolution (1/8º) and are updated at daily frequency. With respects to currents forcing, the Mediterranean Sea model is forced by daily averaged surface currents obtained from CMEMS Med MFC V3 at 1/24º resolution and the North Atlantic model is forced by daily averaged surface currents obtained from the CMEMS Global MFC V2 at 1/12º resolution. These are the CMEMS Global and Med Page 11/ 40

12 MFC surface currents products operational at the time of release of the Med-waves system. A schematic of the Med-waves V3 system is shown in. Also, Table 3 summarizes the Med-waves V3 modelling characteristics. Future upgrades of the Med-waves system will consider coupling to the latest surface currents operational products of CMEMS Global and Med MFC systems. Med-waves generates hourly wave fields over the Mediterranean Sea at 1/24º horizontal resolution. These wave fields correspond either to wave parameters computed by integration of the total wave spectrum or to wave parameters computed using wave spectrum partitioning. In the latter case the complex wave spectrum is partitioned into wind sea, primary and secondary swell. Wind sea is defined as those wave components that are subject to wind forcing while the remaining part of the spectrum is termed swell. Wave components are considered to be subject to wind forcing when where is the phase speed of the wave component, is the friction velocity, is the direction of wave propagation and is the wind direction. As the swell part of the wave spectrum can be made up of different swell systems with quite distinct characteristics it is further partitioned into the two most energetic wave systems, the so called primary and secondary swell. Swell partitioning is done following the method proposed by Gerling (1992) which finds the lowest energy threshold value at which upper parts of the spectrum get disconnected with the process repeated until primary and secondary swell is detected. North Atlantic (Coarse) Model Mediterranean Sea (Fine) WAM Cycle o o Model Domain -75 E 10 W 10oS 70oN oE oW oS oN Horizontal Resolution 1/6o x 1/6o 1/24o x 1/24o Frequency Bins 32 (logarithmically spaced) Hz Direction Bins 24 (equally spaced) Propagation Time-Step Forcing (10m winds) 300sec 60sec 1/8o x 1/8o ECMWF operational analysis & 5-days forecasts Surface Currents CMEMS GLOBAL MFC (1/12o x 1/12o) CMEMS MED MFC (1/24o x 1/24o) Sea Ice 1/8o x 1/8o ECMWF - Table 3: Med-waves V3 modelling characteristics Page 12/ 40

13 Upstream data and boundary condition of the WAM model The CMEMS Med-waves system uses the following upstream data: 1. Atmospheric forcing: NWP 6-h (3-h for the first 3 days of forecast), 1/8º horizontal resolution operational analysis and forecast fields from ECMWF, distributed by the Italian National Meteo Service (USAM/CNMA). 2. Surface currents forcing: GLOBAL_ANALYSIS_FORECAST_PHY_001_024 daily averages at 1/12º (Atlantic model grid forcing) and MEDSEA_ANALYSIS_FORECAST_PHY_006_001 daily averages at 1/24º (Mediterranean model grid forcing) from the CMEMS Global MFC V2 and the CMEMS Mediterranean MFC V3 Analysis and Forecast Systems respectively. 3. Sea-ice cover: daily analysis fields from ECMWF at 1/8º (remain constant during the operational cycle) distributed by the Italian National Meteo Service (USAM/CNMA). Page 13/ 40

14 VALIDATION FRAMEWORK In order to evaluate and assure the quality of the product of the CMEMS Med-waves version V3, the Med-waves system described in Section II has been integrated in hindcast mode for year Model outputs were then compared to observations, focusing on output quality in the Mediterranean Sea. The wave parameters that have been qualified through their comparison with observations include: spectral significant wave height (Hm0) [SWH] spectral moments (0,2) wave period (Tm02) [MWP] The remaining wave parameters included in the CMEMS Med-waves product are not qualified against observational data. This is mainly because there are no relevant observations or these are limited or, as it will be explained below, the observations are of an ambiguous quality. In most of the cases, the quality of these wave parameters is inferred from the quality of those parameters that are thoroughly compared with observations. A valid range, based on climatology and/or expert knowledge, is assigned to each wave parameter. The observations against which modelled wave parameters are compared to include: in-situ observations from moored wave buoys obtained from the CMEMS INSITU_MED_NRT_OBSERV dataset, available through the CMEMS In Situ Thematic Assemble Centre (INS-TAC) satellite altimeter observations from a merged altimeter wave height database setup at CERSAT - IFREMER The INS-TAC is a component of CMEMS which aims at providing a research and operational framework to develop and deliver quality-controlled in situ observations and derived products based on such observations. It is noted however that by the time this report has been written the INS-TAC did not quality-control the wave observations collated from the different providers. As a result, spurious measurements have been identified. These have been confidently filtered though visual inspection for significant wave height, SWH, and mean wave period, MWP, which exhibited a smooth variation. This was not the case for peak wave period, WTPK, which has been found to be particularly spiky and would benefit from an advanced quality-control procedure. This is why, this wave parameter, included in the CMEMS Med-waves product, has not been considered for quality assessment herein despite data availability. Significant wave height measurements from 32 wave buoys within the Mediterranean Sea were available in the examined year. depicts their location and unique ID code. Mean wave period measurements were available from a sub-set of the depicted buoys which excludes all the buoys offshore from the Italian coastline (see Figure 12). To collocate model output and buoy measurements, in space, model output was taken at the grid point nearest to the buoy location. In time, buoy measurements within a time window of ± 1 hr from model output times at 3-hr intervals (0, 3, 6,..., etc) were averaged. Prior to model-buoy collocation, the in-situ observations were filtered so as to remove those values accompanied by a bad quality flag (Quality Flags included in the data files provided by the INS-TAC). After collocation, visual inspection of the data was carried out, which led to some further filtering of spurious data points, as mentioned above. In addition, MWP data below 2 sec Page 14/ 40

15 were omitted from the statistical analysis, since 0.5 Hz (T = 2 sec) is a typical cut-off frequency for wave buoys. Table 5 in section IV includes, amongst others, the number of model-buoy collocation pairs (referred to as "Entries" in the table) obtained at each buoy location and over the entire Mediterranean Sea. Figure 4: Wave buoys' location and unique ID code Satellite observations of significant wave height, SWH, and wind speed, U10 (used to validate the forcing wind fields), for year 2014 were obtained from a merged altimeter wave height database setup at CERSAT - IFREMER (France). This database contains altimeter measurements that have been filtered and corrected (Queffeulou and Croizé-Fillon, 2013). Here, measurements from 3 satellite missions, the JASON-2, CRYOSAT-2, and SARAL, were used. To collocate model output and satellite observations the former were interpolated in time and space to the individual satellite tracks. For each track, corresponding to one satellite pass, along-track pairs of satellite measurements and interpolated model output were averaged over ~50 km (0.5 ) grid cells, centered at grid points of the forcing wind model (0.125 x ). This averaging is intended to break any spatial correlation present in successive 1 Hz (~7 km) observations and/or in neighboring model grid output (Queffeulou, personal communication). Table 3 in section IV includes, amongst others, the number of model-satellite collocation pairs (referred to as "Entries" in the table) obtained over pre-defined sub-regions of the Mediterranean Sea (Figure 1, Section I.3) and over the entire basin. Metrics that are commonly applied to assess numerical model skill and are in alignment with the recommendations of the EU FP7 project MyWave (A pan-european concerted and integrated approach to operational wave modelling and forecasting a complement to GMES MyOcean services, ) have been used to qualify the Med-waves system within the Mediterranean Sea. These include the RMSD, BIAS, Scatter Index (SI), Pearson Correlation Coefficient (CORR), and best-fit Slope (LR_SLOPE). The SI, defined here as the standard deviation of model-observation differences relative to the observed mean, being dimensionless, is more appropriate to evaluate the relative closeness of the model output to the observations at different locations compared with the RMSD which is representative of the size of a typical model-observation difference. The LR_SLOPE corresponds to a best-fit line forced through the origins (zero intercept). In addition to the aforementioned core metrics, merged Density Scatter and Quantile-Quantile (QQ) are provided. The full set of metrics used in the qualification of the Med-waves system is defined in Table 4. Page 15/ 40

16 Name Description Wave parameter Supporting reference dataset Quantity Evaluation of Med-waves using in-situ observations (Full MED) SWH-3H-CLASS2-MOOR-RMSD-MED buoy significant wave height MWP-3H-CLASS2-MOOR-RMSD-MED buoy mean wave period SWH-3H-CLASS2-MOOR-BIAS-MED buoy significant wave height MWP-3H-CLASS2-MOOR-BIAS-MED buoy mean wave period SWH-3H-CLASS2-MOOR-SI-MED buoy significant wave height MWP-3H-CLASS2-MOOR-SI-MED buoy mean wave period SWH-3H-CLASS2-MOOR-CORR-MED buoy significant wave height MWP-3H-CLASS2-MOOR-CORR-MED buoy mean wave period SWH-3H-CLASS2-MOOR-LR_SLOPE-MED buoy significant wave height MWP-3H-CLASS2-MOOR-LR_SLOPE-MED buoy mean wave period 16 Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV RMSD between observations and hindcast, for all Med, for 1-year period and seasonally Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV RMSD between observations and hindcast, for all Med, for 1-year period and seasonally Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV Bias between observations and hindcast, for all Med, for 1-year period and seasonally Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV Bias between observations and hindcast, for all Med, for 1-year period and seasonally Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV SI between observations and hindcast, for all Med, for 1-year period and seasonally Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV SI between observations and hindcast, for all Med, for 1-year period and seasonally Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV Linear correlation between observations and hindcast, for all Med, for 1-year period and seasonally Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV Linear Correlation between observations and hindcast,for all Med, for 1-year period and seasonally Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV Slope of the best fit line between observations and hindcast, for all Med, for 1-year period and seasonally Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV Slope of the best fit line between observations and hindcast,for all Med, for 1-year period and seasonally

17 buoy significant wave height MWP-3H-CLASS2-MOOR-QQ-MED buoy mean wave MWP-3H-CLASS2-MOOR-SCATTER-MED period SWH-3H-CLASS2-MOOR-QQ-MED SWH-3H-CLASS2-MOOR-SCATTER-MED Spectral significant wave height (Hm0) Spectral moments (0,2) wave period (Tm02) Merged Quantile-Quantile and Scatter plots between INSITU_MED_NRT_OBSERV observations and hindcast, for all Med, for 1-year period Merged Quantile-Quantile and Scatter plots between INSITU_MED_NRT_OBSERV observations and hindcast, for all Med, for 1-year period Evaluation of Med-waves using in-situ observations (at buoy locations) SWH-3H-CLASS2-MOOR-RMSDbuoy significant wave <MOORING ID> height MWP-3H-CLASS2-MOOR-RMSDbuoy mean wave <MOORING ID> period SWH-3H-CLASS2-MOOR-BIASbuoy significant wave <MOORING ID> height MWP-3H-CLASS2-MOOR-BIASbuoy mean wave <MOORING ID> period SWH-3H-CLASS2-MOOR-SI-<MOORING buoy significant wave ID> height MWP-3H-CLASS2-MOOR-SI-<MOORING buoy mean wave ID> period SWH-3H-CLASS2-MOOR-CORRbuoy significant wave <MOORING ID> height MWP-3H-CLASS2-MOOR-CORRbuoy mean wave <MOORING ID> period SWH-3H-CLASS2-MOOR-LR_SLOPEbuoy significant wave <MOORING ID> height Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV RMSD between observations and hindcast, for each wave buoy separately, for 1-year period Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV RMSD between observations and hindcast, for each wave buoy separately, for 1-year period Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV Bias between observations and hindcast, for each wave buoy separately, for 1-year period Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV Bias between observations and hindcast, for each wave buoy separately, for 1-year period Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV SI between observations and hindcast, for each wave buoy separately, for 1-year period Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV SI between observations and hindcast, for each wave buoy separately, for 1-year period Spectral significant wave height (Hm0) INSITU_MED_NRT_OBSERV Linear correlation between observations and hindcast, for each wave buoy separately, for 1-year period Spectral moments (0,2) wave period (Tm02) INSITU_MED_NRT_OBSERV Linear correlation between observations and hindcast, for each wave buoy separately, for 1-year period Spectral significant wave height (Hm0) Spectral moments Slope of the best-fit line between observations and INSITU_MED_NRT_OBSERV hindcast, for each wave buoy separately, for 1-year period INSITU_MED_NRT_OBSERV Slope of the best-fit line between observations and Page 17/ 40

18 MWP-3H-CLASS2-MOOR-LR_SLOPE<MOORING ID> SWH-3H-CLASS2-MOOR-QQ-<MOORING ID> SWH-3H-CLASS2-MOOR-SCATTER<MOORING ID> MWP-3H-CLASS2-MOOR-QQ<MOORING ID> MWP-3H-CLASS2-MOOR-SCATTER<MOORING ID> SWH-CLASS4-ALT-RMSD-MED SWH-CLASS4-ALT-BIAS-MED SWH-CLASS4-ALT-SI-MED buoy mean wave period (0,2) wave period (Tm02) hindcast, for each wave buoy separately, for 1-year period buoy significant wave height Spectral significant wave height (Hm0) Merged Quantile-Quantile and Scatter plots between INSITU_MED_NRT_OBSERV observations and hindcast, for each wave buoy separately, for 1-year period buoy mean wave period Spectral moments (0,2) wave period (Tm02) Merged Quantile-Quantile and Scatter plots between INSITU_MED_NRT_OBSERV observations and hindcast, for each wave buoy separately, for 1-year period Evaluation of Med-waves using satellite observations (full MED) Comparison to Merged altimeter wave Spectral significant RMSD between observations and hindcast, for all Med, altimeter significant height database from wave height (Hm0) for 1-year period and seasonally wave height CERSAT - IFREMER Comparison to Merged altimeter wave Spectral significant Bias between observations and hindcast, for all Med, altimeter significant height database from wave height (Hm0) for 1-year period and seasonally wave height CERSAT - IFREMER Comparison to Merged altimeter wave Spectral significant SI between observations and hindcast, for all Med, for altimeter significant height database from wave height (Hm0) 1-year period and seasonally wave height CERSAT - IFREMER Page 18/ 40

19 SWH-CLASS4-ALT-CORR-MED Comparison to altimeter Spectral significant significant wave height wave height (Hm0) SWH-CLASS4-ALT-LR_SLOPE-MED Comparison to altimeter Spectral significant significant wave height wave height (Hm0) SWH-CLASS4-ALT-QQ-MED SWH-CLASS4-ALT-SCATTER-MED Comparison to altimeter Spectral significant significant wave height wave height (Hm0) Merged altimeter wave height database from CERSAT - IFREMER Merged altimeter wave height database from CERSAT - IFREMER Merged altimeter wave height database from CERSAT - IFREMER Linear correlation between observations and hindcast, for all Med, for 1-year period and seasonally Slope of the best fit line between observations and hindcast, for all Med, for 1-year period and seasonally Merged Quantile-Quantile and Scatter plots between observations and hindcast, for all Med, for 1-year period Evaluation of Med-waves using satellite observations (MED sub-regions) SWH-CLASS4-ALT-RMSD-<REGION> Comparison to altimeter Spectral significant significant wave height wave height (Hm0) SWH-CLASS4-ALT-BIAS-<REGION> Comparison to altimeter Spectral significant significant wave height wave height (Hm0) SWH-CLASS4-ALT-SI-<REGION> Comparison to altimeter Spectral significant significant wave height wave height (Hm0) SWH-CLASS4-ALT-CORR-<REGION> Comparison to altimeter Spectral significant significant wave height wave height (Hm0) SWH-CLASS4-ALT-LR_SLOPE-<REGION> Comparison to altimeter Spectral significant significant wave height wave height (Hm0) Merged altimeter wave height database from CERSAT - IFREMER Merged altimeter wave height database from CERSAT - IFREMER Merged altimeter wave height database from CERSAT - IFREMER Merged altimeter wave height database from CERSAT - IFREMER Merged altimeter wave height database from CERSAT - IFREMER Table 4: List of metrics for Med-waves evaluation using in-situ and satellite observations Page 19/ 40 RMSD between observations and hindcast, at Med subbasins, for 1-year period Bias between observations and hindcast, at Med subbasins, for 1-year period SI between observations and hindcast, at Med subbasins, for 1-year period Linear correlation between observations and hindcast, at Med sub-basins, for 1-year period Slope of the best fit line between observations and hindcast, at Med sub-basins, for 1-year period

20 VALIDATION RESULTS Significant wave height Comparison with in-situ observations Table 5 shows results of the comparison between hindcast SWH (model data) and in-situ observations (reference data), for the Mediterranean Sea as a whole, for the entire year of 2014 and seasonally. In the table, "Entries" refers to the number of model-buoy collocation pairs, i.e. to the sample size available for the computation of the relevant statistics, is the mean reference value, is the mean model value, STD R and STD M are the standard deviations of the reference and model data respectively. The remaining quantities are the qualification metrics defined in the previous section. is the respective merged QQ-Scatter plot for the full 1-year period. In the figure, the QQ-plot is depicted with black crosses. Also shown are the best fit line forced through the origin (red solid line) and the 45 reference line (red dashed line). MED ENTRIES STD R STD M RMSD SI BIAS CORR LR_SLOP E Whole Year Winter Spring Summer Autumn Table 5: Med-waves SWH evaluation against wave buoys' SWH, for the full Mediterranean Sea, for 1 year period (2014) and seasonally. Relevant metrics from Table 2: SWH-3H-CLASS2-MOOR_xxxx-MED (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE) Table 5 shows that the typical difference (RMSD) varies from 0.17 m in summer to 0.25 m in winter. However, the scatter in summer (0.26) is about 2% higher than the scatter in winter (0.24) whilst a lower correlation coefficient is associated with the former season. This suggests that the model follows better the observations in 'stormy' conditions, with well-defined patterns and higher waves. A similar conclusion has been derived by other studies (Cavaleri and Sclavo, 2006; Ardhuin et al., 2007; Bertot et al., 2013) with respects to wind and wave modelling performance in the Mediterranean Sea. Alike summer, autumn is characterized by lower mean wave height, higher scatter and lower correlation coefficient compared to winter and spring. Small negative BIAS with a variation of below 3% between seasons is observed, with the highest relative bias (BIAS/ ) obtained for autumn ( 5%) and the smallest for spring ( 2.5%). LR_SLOPES are also below unity (-) with small variation between seasons. These values are indicative of an underestimation of the wave height in the Mediterranean Sea by the model, a result which is in agreement with the results of a number of operational or pre-operational models for the Mediterranean Sea (e.g. JCOMM, 2014; Donatini et al., 2015) and is most probably linked to an underestimation of the wind speed by the ECMWF forcing wind model (see ). Overall, the spring statistics are the ones closest to the year-long statistics for the Mediterranean Sea. 20

21 RMSD Figure 5: QQ-Scatter plots of Med-waves output SWH (Hs) versus wave buoys' observations, for the full Mediterranean Sea, for 1 year (2014) period: QQ-plot (black crosses), 45 reference line (dashed red line), least-squares best fit line (red line). Relevant metrics from Table 2: SWH-3H-CLASS2-MOOR-QQMED, SWH-3H-CLASS2-MOOR-SCATTER-MED depicts the pattern of the agreement between hindcast and observed SWH for different SWH value ranges. The figure reveals that the SWH underestimation by the model is mainly occurring for wave heights below 4 m and is rather small. It also shows a QQ-plot that is really close to the reference line over most of the SWH range observed, which means that waves of a specific wave height have a very similar probability of occurrence in the hindcast and in the observations. 'Outliers' are present in the scatter plot, for example, a number of measured waves of m height are not simulated by the model (not enough evidence was found to remove the depicted outliers from the calculation of the statistics as faulty). These values are likely to have introduced some bias to the computed statistics and, clearly, have led to a greater SI. Table 6 shows results of the comparison between hindcast SWH and in-situ observations for each of the wave buoys depicted in. The qualification metrics for the different buoy locations (from east to west) are also plotted in in order to facilitate the visualization and interpretation of the relative performance of the wave model at the different locations. To be able to readily compare the pattern of variation of the different metrics at the different locations the absolute BIAS and the CORR and LR_SLOPE deviations from unity are plotted in the bottom plot of. The values of these metrics as given in Table 6 are shown in the middle plot. For convenience, a map of the wave buoy locations is included in the figure (top). Table 6 reveals that the typical difference (RMSD) at the different buoy locations varies from 0.16 m to 0.44 m with the highest values (> 0.3 m) observed in the North Adriatic Sea at buoy location (0.44) and offshore the French coastline at buoy location (0.35). Both locations are associated with poor overall qualification metrics with location in particular displaying the poorest overall statistics. Thus, SI, which varies between 0.17 (61197) and 0.53 (61218), also obtains its highest values indicating a poorer model performance at locations within the North Adriatic Sea. Locations SARON and follow with values of 0.4 and 0.32 respectively. In general, SI values above the mean value for the whole Mediterranean Sea (0.25) are obtained at all wave buoys located near the coast that are Page 21/ 40

22 STD R STD M RMSD SI BIAS CORR LR_SLOPE SARON ATHOS Buoy ID ENTRIES Table 6: Med-waves SWH evaluation against wave buoys' SWH, for each individual buoy location, for 1 year period (2014). Relevant metrics from Table 2: SWH-3H-CLASS2-MOOR-xxxx-<MOORING ID> (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE and "<MOORING ID"> can be replaced with the moored buoy ID of the relevant wave buoy) Page 22/ 40

23 LR_SLOPE RMSD LR_SLOPE deviation Figure 6: SWH metrics (middle, bottom) at buoy locations (top) for 1 year period (2014) (plots display metrics starting from the western buoy location and moving eastwards. Last value corresponds to the full Mediterranean Sea). Bottom plot: CORR and LR_SLOPE deviations from unity are shown. sheltered by land masses on their west (e.g. western French coastline, eastern part of Corsica and eastern part of Italy). With respects to the Italian buoys, this result is in close agreement with Cavaleri and Sclavo (2006) who found that the performance of the operational ECMWF wave model deteriorated at those Italian wave buoys facing East. This is because the resolution of the forcing wind model is not capable of well reproducing the fine interaction between the prevailing northpage 23/ 40

24 northwesterly winds in the northern Mediterranean Sea with the complex orography sheltering the northern Mediterranean coastline. An underestimation of wind speeds and consequently of wave heights (also the case herein) is commonly observed at such locations (Ardhuin et al., 2007). In addition, the buoys are often located only a few kilometers from the coastline, thus, in these conditions, i.e. when the wind is blowing from the coast, the approximation of the wave model grid size can lead to non-negligible fetch differences. Similar SI values are found within enclosed basins characterized by a complex topography such as the Adriatic and Aegean Seas. In general, the more close the location to the coastline (e.g ) and/or the more complex the surrounding topography (e.g. SARON) the poorer the model performance expected. As explained in several studies (e.g. Cavaleri and Sclavo, 2006; Bertot et al., 2013; Zacharioudaki et al., 2015), in these cases, the spatial resolution of the wave model is not adequate to resolve the fine bathymetric features whilst, as mentioned above, the spatial resolution of the forcing wind model is incapable to reproduce the fine orographic effects, introducing errors to the wave hindcast. The correlation coefficient (CORR) largely follows the pattern of variation of the SI. It ranges from 0.79 (61218) in the Adriatic Sea to 0.98 (61213) at a location west of Sardinia which is well exposed to the prevailing north-westerly winds in the region. BIAS varies from m at location in the Adriatic Sea to 0.11 m at location offshore from the French coastline. It is mostly negative indicating an underestimation of the observed wave height by the model, with positive BIAS observed at only 6 out of the 32 buoy locations examined. In most of the cases of relatively high positive BIAS, this is because the wave model resolution has missed or has partly captured important bathymetric features in the surroundings of the relevant locations, thus, missing or reducing the shadowing effects produced by these features. For example, shoreward buoy where the largest positive BIAS is observed there are few small islands that are almost entirely missed by the model. Also, at buoy 61221, similarly to Bertot et al. (2013), the coastal geometry is not well represented by WAM. Bertot et al. (2013) state that the buoy position is exposed to the easterly waves more than is actually the case, leading to the observed overestimate by the model. Similar conclusions stand for SARON buoy in the Aegean Sea. Overestimation at buoy in the Alboran Sea is part of a general overestimation of the wave heights in the Atlantic and Alboran regions as it will be seen later in the comparison with the satellites. In accordance with BIAS, best-fit slopes (LR_SLOPE) vary from 0.77 at buoy location in the Adriatic Sea to at buoy location offshore from France. LR_SLOPE values above unity coincide with locations of positive BIAS, otherwise they are below unity, confirming an overall underestimation of the observations by the model. Relatively low LR_SLOPE values of below 0.9 are confined at buoy locations in the Italian Seas and location offshore from the eastern part of the French coastline. In general, the pattern of variation of LR_SLOPE is close to the pattern of variation of BIAS. Up to now, the overall performance of the Med-waves modelling system at the different wave buoy locations has been analyzed independently of the severity of the conditions. Error: Reference source not found similarly to Figure 5, shows the QQ-Scattter plots of hindcast SWH versus measured SWH at four buoy locations, exhibiting model performance over the different wave height ranges. The results at these four locations are reasonably representative of the different behaviours of the wave model at the different wave buoy locations in the Mediterranean Sea shown in. Thus, the top left plot shows the behaviour of the model at location 61188, offshore from the border between France and Spain, backed on the west by the Pyrenees Mountains. It is seen that model underestimation occurs throughout the measured SWH range except from the highest percentiles of SWH where model overestimation is observed. This distribution, with a smaller or larger model underestimate and with a more or less pronounced convergence or overestimate towards the highest waves, is observed at the majority of the wave buoy locations. The top right plot shows the scatter at location in the Adriatic Sea where the model has its poorest performance. As it is apparent from the plot the model has failed to Page 24/ 40

25 reproduce a number of stormy wave events at the location. Clearly, these 'outliers' (also in ) introduce some bias to the computed statistics. Time-series plots (not shown) revealed that these events happened between the 10th and 15th of January 2014 and mostly belong to a single storm. In the future, a more thorough investigation of model output and observations during this period, including wave directions, may help explain the reasons for this model failure. The bottom left plot corresponds to location 61221, south of the island of Sardinia. There, the model overestimates the observed SWH over the entire SWH range, even more so in the upper end of this range. Considerable model overestimation, mostly over the middle and higher SWH ranges, is observed at all wave buoys associated with unresolved bathymetric features in their surroundings (e.g , SARON). Over the lower SWH range, converge or underestimate is also observed in these conditions. At SARON (not shown), the surrounding topography is highly complex, including both RMSD RMSD RMSD RMSD Figure 7: QQ-Scatter plots of Med-waves output SWH (Hs) versus wave buoy observations at specific wave buoy locations, for 1 year (2014) period: QQ-plot (black crosses), 45 reference line (dashed red line), least-squares best fit line (red line). Relevant metrics from Table 2: SWH-3H-CLASS2-MOOR-QQ<MOORING ID> (where "<MOORING ID>" can be replaced with the moored buoy ID of the relevant wave buoy) orographic and bathymetric effects, resulting in highly scattered data around the reference line. The bottom right plot shows results at buoy location east of the Balearic Islands. This is a wellexposed offshore wave buoy; consequently, the behaviour of the model at this location is expected to be representative of its performance at well-exposed offshore sites. A relatively small scatter of the data points is shown in the plot with QQ crosses and best-fit line laying close to the reference line. Page 25/ 40

26 More specifically, the model converges to the observations for wave heights below about 2 m, somewhat underestimates the observations for wave heights between 2-4 m, and tends to overestimate SWH for higher waves. A very similar distribution is found for location west of the Balearic Islands. Other well-exposed offshore locations present QQ-Scatter patterns that are not far from the one shown for location Page 26/ 40

27 Comparison with satellite observations This sub-section starts with the comparison of Med-waves hindcast SWH with satellite observations of SWH separately for each satellite. This is done for 1-year period (2014) for the full Mediterranean Sea and for the different sub-regions defined in Figure 1. Respective results are shown in Table 7 and Figure 8. Table 7 shows that even though the model-satellite comparison behaves similarly for the three different satellites in terms of SI and CORR, a substantially more negative model BIAS associated with a considerably lower LR_SLOPE is found for CRYOSAT-2. Absolute model BIAS is more than 10% higher when the model is compared to CRYOSAT-2. RMSD is also higher. Figure 8 shows that these results are largely consistent between the different Mediterranean sub-regions although they are more pronounced in the Western Mediterranean and the Adriatic Sea. A lower model underestimate of the CRYOSAT-2 measurements is observed in the Ionian Sea and the Eastern Mediterranean. The statistics of model-jason-2 and model-saral comparisons are comparable, with the model exhibiting its best performance when compared to the observations of SARAL. Satellite JASON-2 SARAL CRYOSAT-2 ENTRIES STD R STD M RMSD SI BIAS CORR LR_SLOP E Table 7: Med-waves SWH evaluation against satellite SWH, for the full Mediterranean Sea, for 1 year period (2014). Relevant metrics from Table 4: SWH-CLASS4-ALT_xxxx_MED (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE) RMSD LR_SLOPE Fi gure 8: Med-waves SWH evaluation against satellite SWH, for each Mediterranean Sea sub-region shown in Figure 1, for 1 year period (2014). Relevant metrics from Table 4: SWH-CLASS4-ALTxxxx_<REGION> (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE and Page 27/ 40

28 "<REGION>" can be replaced with the relevant sub-region name) It was decided to exclude the observations of CRYOSAT-2 from the analysis. Apart from the aforementioned discrepancies, there are other results in the literature to support this decision. Specifically, satellite-buoy comparisons performed by Sepulveda et al. (2015) have shown that SARAL SWH is of better quality than JASON-2 and CRYOSAT-2 SWH at both open ocean and coastal buoy sites. In fact, SARAL data are of very high quality with no need of corrections whilst corrections are applied to JASON-2 and CRYOSAT-2 SWH observations (corrected data are used herein). After corrections, JASON2 SWH has been found to well approximate SARAL SWH whilst less accurate results have been obtained for CRYOSAT-2, particularly for wave heights below 1.5 m. For these reasons, in what follows, the comparison of Med-waves SWH is performed against merged satellite observations from SARAL and JASON-2 satellites which are of similar accuracy. Table 8 shows statistics from the comparison of the Med-waves hindcast SWH and satellite observations of SWH, for the full Mediterranean Sea, for 1-year period and seasonally. (right) shows the corresponding QQ-Scatter plot for 1-year period, for the full Mediterranean Sea. Figure 9 (left) shows an equivalent QQ-Scatter plot resulting from the comparison of the ECMWF forcing wind speeds, U10, and JASON-2 measurements of U10 (no U10 available from SARAL or CRYOSAT-2). MED ENTRIES STD R Whole Year Winter Spring Summer Autumn STD M 0.75 RMSD SI BIAS CORR LR_SLOP E Table 8: Med-waves SWH evaluation against satellite SWH (JASON-2 and SARAL), for the full Mediterranean Sea, for 1 year period (2014) and seasonally. Relevant metrics from Table 2: SWHCLASS4-ALT-xxxx-MED (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE) (left) shows that the ECMWF forcing wind model underestimates observed U10 throughout the entire U10 range, even more so at high wind speeds. An overall model underestimation of 9% associated with a LR_SLOPE of 0.9 have been computed. Figure 9 (right) also shows an overall Med-waves model underestimation of observed SWH by about 5% associated with a LR_SLOPE of. Nevertheless, in this case, the model somewhat underestimates observed SWH over the lower SWH range (< 2 m), converges to the observed SWH over the middle SWH range (2-3.5 m) while, generally, somewhat overestimates the larger waves in the data records. This apparent discrepancy between wind and wave scatter distributions is a consequence of the modification of the default values of the whitecapping dissipation coefficients in WAM (see Section II.2). A QQ-Scatter obtained before this modification (not shown) is indeed very similar to the one of the ECMWF wind speeds in. On the whole, Figure 9 shows that the performance of Med-waves at offshore locations in the Mediterranean Sea (satellite records near the coast are mostly filtered out as unreliable) is very good. Comparing to the equivalent results obtained from the model-buoy comparison (), a very similar pattern of scatter distribution is observed in the two plots, also evident from the orientation of the best-fit lines and the curvature of the QQplots. A smaller scatter (by about 6%) with a larger overall bias (by about 2%) is associated with the Page 28/ 40

29 model-satellite comparison. SI values compare well at the more exposed wave buoys in the Mediterranean Sea. Table 7 shows the seasonal variation of the Med-waves model performance. The typical difference (RMSD) varies from 0.17 m in summer to 0.24 m in winter. SI is highest in summer (0.2) and lowest in winter (0.17). Its pattern of variation is similar to the one obtained from the model-buoy comparison despite the lower SI values (by 6-8%) in the current case. The main difference is that, in the former case, the highest SI value has been computed for autumn. Correlation coefficient varies accordingly. In general, as explained in the previous sub-section, a lower scatter with a higher correlation is expected the more well-defined the weather conditions are. Alike in the model-buoy comparison, BIAS is negative in all seasons. Its highest relative value (BIAS/ ) of 7.7% is computed for autumn and its lowest of 3.5% for summer. LR_SLOPE varies from in spring and autumn to 0.97 in winter. Overall, Table 7, alike Table 5, reveals that the statistics of spring are the most representative of the year-long statistics for the Mediterranean Sea. RMSD RMSD Figure 9: QQ-Scatter plots of: (left) ECMWF forcing wind speed U10 versus satellite U10 (JASON-2); (right) Med-waves SWH (Hs) versus satellite SWH (JASON-2 and SARAL), for the full Mediterranean Sea, for 1 year (2014) period. Relevant metrics from Table 2: SWH-CLASS4-ALT-QQ-MED and SWH-CLASS4ALT-SCATTER-MED Table 9 shows the statistics of the comparison of the Med-waves hindcast SWH and satellite observations of SWH for the different sub-regions of the Mediterranean Sea defined in Figure 1. For visualization purposes, Figure 10 (right column) maps the statistics shown in Table 9. In addition, equivalent statistics are mapped for the ECMWF - satellite comparison of wind speeds (left column). It is noted that the Relative BIAS (BIAS/ ) is displayed in the figure. This quantity allows for a more straightforward comparison between the different sub-basins in terms of percentage deviations from the observed mean value. It is also highlighted that the spatial coverage of the model-satellite wind collocations (measurements only from JASON-2) is much more limited than the spatial coverage of the model-satellite wave collocations (measurements from both SARAL and JASON-2). As a consequence, the wave statistics are expected to be more representative of the sub-regions under consideration compared to the wind statistics. This is particularly true for the Adriatic, the Ligurian and the Alboran Seas. In addition, the wave statistics have been computed using a sample size of at least 400 data points whilst the wind statistics have been obtained with a minimum sample requirement of 200 data Page 29/ 40

30 points. Thus, the confidence associated with the wave statistics is higher than the confidence associated with the wind statistics. For the above reasons, the wind metrics presented in Figure 10 are interpreted with caution. Figure 10 (right column) shows that the typical difference (RMSD) varies from 0.18 m in the SouthCentral Levantine Basin (lev3 in Figure 1) to 0.29 m in the Atlantic sub-region (atl). In the Mediterranean Sea, relatively higher RMSD ( m) is observed in the Alboran Sea (alb), the North-West Mediterranean (nwm), the Adriatic (adr1, adr2) and the Aegean (aeg) Seas. Relatively lower RMSD is Satellite ENTRIES STD R STD M RMSD SI Atl Alb swm swm Nwm tyr tyr ion ion ion adr1 809 adr2 655 lev1 lev2 BIAS CORR LR_SLOPE lev lev Aeg Table 9: Med-waves SWH evaluation against satellite SWH (JASON-2 and SARAL), for each individual Mediterranean Sea sub-region shown in Figure 1, for 1 year period (2014). Relevant metrics from Table 2: SWH_CLASS4_SAT_xxxx_BASIN (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE and "BASIN" can be replaced with the relevant sub-region name) found for the southern sub-regions of the Mediterranean Sea, particularly over the eastern part of the basin. Relatively higher values of SI are also found in the Alboran, Adriatic and Aegean Seas. The Tyrrhenian Sea, mainly its northern part (tyr1), and the East Levantine Basin (lev4) follow. SI and CORR have a similar pattern of variation; a notable difference being that the correlation coefficient obtains its worst value in the East Levantine and, in general, it has relatively lower values in the well-exposed regions of the Levantine Basin compared to the well-exposed regions on its west. In accordance with the above results, Ratsimandresy et al. (2008), examining model-satellite agreement over coastal locations of the western Mediterranean Sea, found the worst correlations in the Alboran Sea and east of Corsica Island. Bertot et al. (2013), in a comparison of high resolution wind and wave model output with satellite data over different sub-regions of the Mediterranean Sea, also found the largest scatter and lowest correlations in the Adriatic and the Aegean Seas. In agreement, Zacharioudaki et al. (2015), focusing on the Greek Seas, have shown a considerably larger scatter in the Aegean Sea than in the surrounding seas, when model output was compared to satellite observations. As explained in the previous sub-section (model-buoy comparison), it is the difficulties of wind models to well-reproduce Page 30/ 40

31 orographic effects and/or local sea breezes and the difficulties of wave models to well-resolve complicated bathymetry that introduce errors in these fetch-limited, enclosed regions, often characterized by a complex topography. Indeed, comparison with the equivalent results for the ECMWF wind speeds confirms these difficulties. For example, the pattern of SI and CORR variation for U10 largely resembles that for SWH, corroborating the conclusion of many studies that errors in wave height simulations by sophisticated wave models are mainly caused by errors in the generating wind fields (e.g. Komen et al., 1994; Ardhuin et al., 2007). Nevertheless, some differences do exist. For instance, the SWH SI in the Aegean Sea is relatively higher than the corresponding U10 SI. This is most probably because in this region of highly complicated bathymetry with many little islands the error of the wave model increases in relation to the error of the wind model. Similarly, in the East Levantine, SWH SI is lower than that implied by U10 SI. In this case, the wind model may not well simulate local wind patterns, characterized by local sea breezes and easterly directions (Galil et al., 2006), however, the wave regime which is dominated by waves from the west sector (Galil et al., 2006) is better reproduced by the wave model. Negative BIAS and LR_SLOPE below unity are the case in all subregions except for the Atlantic and Alboran Sea. In the latter regions, the wave model overestimates the observations by 5-6%. Otherwise, it underestimates the observations by about 2% in the East South-West Mediterranean (swm1) and Aegean Seas to about 15% in the Adriatic (adr2). In general, the largest biases are found in the Adriatic (adr1, adr2), the North Ionian Sea (ion3) and the Levantine Basin (lev2,lev3,lev4) with values of %. LR_SLOPE varies accordingly with values between 0.88 in the South Adriatic Sea (adr2) and 0.99 in the Aegean Sea and up to 1.05 in the Alboran Sea. A notable difference between Relative BIAS and LR_SLOPE distribution is seen in the North Adriatic Sea (adr1). There, a lower LR_SLOPE would have been expected considering the value of the Relative BIAS. Looking at the associated QQ-Scatter plots (not shown), an underestimation of waves below about 1.5 m together with an overestimation of higher wave heights is observed. This scatter distribution reduces the discrepancies between the best-fit line and the reference line when the former is forced through the origin. If not, a best-fit line with a slope close to or above unity and a negative intercept would have been obtained. In general, the QQ-Scatter plots (not shown) have revealed that, in most of the cases, the model underestimates the observations over the lower SWH range but tends to converge or even overestimate the observations as SWH increases. For LR_SLOPE below unity, the larger the LR_SLOPE the fastest the convergence and overestimation, for LR_SLOPE above unity, the overestimation is widespread. Comparing with the equivalent results for the ECMWF wind speed, it is evident that although there are similarities in the Relative Bias and LR_SLOPE distributions, there are also considerable differences. In general, in terms of absolute value, the Relative Bias associated with the wind field is larger than that associated with the wave field except for the South Adriatic Sea, the Alboran Sea and the Atlantic. In fact, in the latter two regions, a change of sign from negative to positive is observed between wind and waves. As already mentioned, this is a consequence of the modification of the whitecapping dissipation coefficients from default values in WAM, which has led to an important offset of the negative BIAS associated with the ECMWF wind speeds, especially over the high SWH range. Thus, in regions where the ECMWF underestimate has been small, as in the Atlantic, modification of the dissipation coefficients has eventually led to an overshoot of the observed SWH. This is a robust pattern obtained for the whole Atlantic area simulated by the nested Med-waves model (up to W, ). Figure 10 shows that the modification of the whitecapping dissipation coefficients has particularly benefited the western Mediterranean Sea, west of the islands of Sardinia and Corsica. The increase in negative SWH Relative BIAS in the South Adriatic Sea relative to the respective U10 Relative BIAS is somewhat puzzling, however, as mentioned above, small confidence is pertained to the results of U10 in the Adriatic Sea due to a limited observational coverage by JASON-2. Page 31/ 40

32 RMSD (m/s) RMSD SI SI Relative BIAS Relative BIAS CORR CORR LR_SLOPE LR_SLOPE Figure 10: ECMWF U10 (left column) and Med-waves SWH (right column) evaluation against satellite U10 (JASON-2) and satellite SWH (JASON-2 and SARAL) respectively: maps of metric values over the Mediterranean Sea sub-regions shown in Figure 1, for 1 year period (2014). Page 32/ 40

33 Mean Wave Period Comparison with in-situ observations RMSD Figure 11: QQ-Scatter plots of Med-waves output MWP (Tm) versus wave buoys' observations, for the full Mediterranean Sea, for 1 year (2014) period. Relevant metrics from Table 2: MWP-3H-CLASS2MOOR-QQ-MED and MWP-3H-CLASS2-MOOR-SCATTER-MED Table presents the statistics of the comparison between the Med-waves hindcast mean wave period, MWP, and in-situ observations of mean wave period, for the full Mediterranean Sea, for 1 year period (2014) and seasonally. Figure 11 shows the corresponding QQ-Scatter plot for the year-long statistics. It is shown that the model exhibits greater variability than the observations (STD in Table ). RMSD varies from 0.8 s in summer to 1.07 s in winter. In relation to the mean of the observations, the modelobservation difference is about 17-19%, with winter and spring being at the low end of this range and autumn at the high. SI has a small variation with values of 0.12 in winter and spring and 0.14 in summer and autumn. The non-trivial deviation of SI from relative RMSD (RMSD/ ) indicates that a substantial part of the model-observation difference is caused by BIAS. CORR has its minimum value (0.78) in summer and its maximum (0.87) in winter and spring. As before, these results indicate that the model wave period, alike the model wave height, better follows the observations in well-defined wave conditions of higher waves and larger periods. BIAS is negative with values that correspond to a model underestimate of about %. Correspondingly, LR_SLOPE has a small variation of Like for wave height, spring statistics are the most representative of the year-long statistics. Figure 11 clearly shows that the wave model underestimates the observations throughout the observed MWP range. Measurements of MWP < 4.5 s are especially underestimated while those of relatively high MWP are better approximated by the model. STD R STD M RMSD SI BIAS CORR LR_SLOP E Buoy ID ENTRIES Page 33/ 40

34 SARON ATHOS Table 10: Med-waves MWP evaluation against wave buoys' MWP, for each individual buoy location, for 1 year period (2014). Relevant metrics from Table 2: MWP-3H-CLASS2-MOOR-xxxx-<MOORING ID> (where "xxxx" represents each of the metrics RMSD, SI, BIAS, CORR, LR_SLOPE and "<MOORING ID>" can be replaced with the moored buoy ID of the relevant wave buoy) Table 10 gives the statistics of the model-buoy comparison at the individual wave buoy locations presented in except for the wave buoys offshore from the Italian coastline (see )., alike, plots those statistics to facilitate their visualization and interpretation. As in, the absolute BIAS and the deviations of CORR and LR_SLOPE from unity are plotted. In this case the values of BIAS and LR_SLOPE as given in Table 10 are not included in the figure because BIAS is always negative and LR_SLOPE always below unity. The typical model-observation difference relative to the mean of the observations (RMSD/ ) has its lowest values of 12-16% over the western part of the Mediterranean Sea, west and south of France, with the two locations nearest to the Gibraltar Straight being at the low end of this range. Otherwise, this difference is 17-23% reaching up to 29% at location near the French-Italian border. At this location, all qualification metrics obtain their worst value. This is because wave buoy is located at a distance below 2 km from coast and is affected by winds blowing from land. As already explained in the model-buoy wave height comparison, in this situation, the simulated fetch may differ substantially from the actual fetch because of the wave model grid size approximation; moreover, wind speed and wave height is considerably underestimated. In, it is evident that the RMSD is mainly caused by BIAS which is negative at all locations. Thus, accordingly to the RMSD, the relative BIAS is below 10% over the western Mediterranean, is 14%-20% otherwise reaching up to 23% at location It is only at Page 34/ 40

35 LR_SLOPE deviation Figure 12: MWP metrics (bottom) at buoy locations (top) for 1 year period (2014) (plots display metrics starting from the western buoy location and moving eastwards. Last value corresponds to the full Mediterranean Sea). CORR and LR_SLOPE deviations from unity are shown. location 61197, offshore from the eastern Balearic islands, that the scatter of the data appears to contribute more to the typical model-observation difference than the BIAS. At this well-exposed offshore location where model performance is optimal for both wave height and wave period the relative BIAS is only 2% (lowest value) whilst the associated LR_SLOPE is 0.97 (highest value). The relatively high SI (0.15) and moderate correlation (0.86) at this location could be associated with the appearance of two density peaks in the density Scatter plot (not shown), indicative of a double peaked frequency spectrum. Density scatter plots with two peaks, although less distinct, have also been obtained for locations and 61021, offshore from France. In general, a close examination of the QQ-Scatter plots (not shown) corresponding to the different locations has revealed that the model largely underestimates the observed MWP over the lower wave period range at all locations. Over the higher range, the model overestimates the observed MWP in the western Mediterranean Sea, west and south of France. Otherwise, model underestimates all observed MWP with some convergence towards higher values. SI is relatively small with values between 0.09 (ATHOS) and 0.18 (61187) while CORR varies from 0.65 (61187) to 0.9 (61430, 68422). Generally, similarly to the wave height results, Page 35/ 40