Figure 4.16: Comparison between the calibration run with SATKDT=default + 0.9 x 10-6 and REFKDT=0.3 (black line) and the streamflow observations (red line) at the Monte Molino river section.(m3/sec) .
In order to investigate more the model calibration results, the analysis is carried on considering not only the streamflow values observed in a point section, but also the distributed fluxes variables and soil moisture fields. The results are shown in the following section (section 4.4.5).
4.4.4 Flux evaluation
Fluxes comparison are evaluated between the best WRF-Hydro model calibration run (REFKDT= 0.3 and SATKDT= default + 0.9x10-6) and the available observations of the flux stations for the year 2012 (Fig. 4.5). In particular the distribution between latent and sensible heat and the net radiation is analyzed. Informations of the model are extrapolated at the nearest pixel to the observation site. Latent and sensible heat fluxes are aggregated at the daily scale, both for model output and station observations. In the scatterplot of Fig. 4.18 are shown the results for the ITCA1 station in terms of sensible and latent heat partitioning for model and observations.
Figure 4.18: Daily latent and sensible heat partitioning between WRF-Hydro best cali- bration run (black) and observations (blue) at the ITCA1 station site, with the associated statistics in terms of RMSE, R2 and regression coefficient.
Fig.4.19 displays the hourly values scatterplot between observed and modelled net radiation for the same station. The complete set of figures for all the station are shown in the appendix C.1 and appendix C.2.
The graphs are evaluated considering the linear regression coefficient for latent and
Figure 4.19: Net radiation scatterplot between WRF-Hydro best calibration run and observations at the ITCA1 station site, with the associated statistics in terms of RMSE, R2 and regression coefficient.
sensible heat fluxes and the RMSE, coefficient of determination (R2) and linear regression coefficient for the net radiation. The quantitative results are displayed in Tab 4.4.
The sensible and latent heat partitioning seems to differ between model and obser- vations. In particular the sensible heat flux in the model presents also negative values, while they are never observed in the station. At the contrary, the modelled latent heat never goes under zero, while observations do. In addition to that, the linear regression coefficients have values that are quite different between model and observations and the scatterplots show different distributions. In the evaluation, the geographical positioning of the stations in the area has to be taken into account. As stated in section 4.2.2, the flux station are situated outside the Tiber river basin, where actually the model is cali- brated and they are not well spatially distributed in the domain. Even if these stations are the only available in the domain area and represent a valuable contribution to the comparison, the results of these analysis has to be handled with care because of the above mentioned reasons. Moreover, even if the flux station data used are the highest data quality possible (level 3 quality dataset in the Fluxnet repository), some biases in the observation cannot be excluded.
At the contrary of what observed in terms of sensible and latent heat, the net radiation shows a good accordance between model and observations. Points are not too dispersed in the scatterplot results, correlation coefficient are very close to one (range 0.92-1.00) in
Table 4.4: Quantitative analysis of flux comparison of sensible and latent heat partition- ing and net radiation
Latent and sensible heat Net Radiation Stations ObsVegType LUIndexHydro RegrHydro RegrObs RMSE Corr RegCoeff IT CA1 Broadleaf
Forest Mix
shrub-grassland 0.29 0.03 23.32 0.99 1.02
IT CA2 Cropland Mix
shrub-grassland 0.29 0.49 20.23 0.99 1.10
IT CA3 Broadleaf
Forest Mix
shrub-grassland 0.29 0.35 9.09 1.00 1.05
IT RO4 Cropland Cropland 0.45 0.13 4.05 1.00 1.01
IT COL Broadleaf
Forest Dry Cropland-
Pasture 0.25 0.35 76.85 0.92 0.73
all the stations and RMSE values are very low. Also the variability of the model is very similar to the observed variability, as stated by the R2 coefficients.
Finally, an additional check is performed over the associated land use type in the station site. The vegetation type of the station site are compared with the corresponding vegetation type of the observations, in order to check if the local characteristics cor- respond to the modelled ones. Only the ITRO4 station shows accordance in term of vegetation type. If on one hand this same land use feature does not help in a better representation of the latent and sensible heat partitioning for that specific station, on the other hand the net radiation results are the the best for all the scores at ITRO4 station site.
4.4.5 Soil moisture evaluation
Soil moisture evaluation has been carried on comparing the results of the best calibration model run (REFKDT= 0.3 and SATKDT= default + 0.9x10-6) with the available soil moisture station measurements in the catchment (Fig 4.4). The WRF-hydro model output are compared with the soil moisture observations, extracting the nearest model pixel to the station. For every extracted pixel, an additional check on the slope, soil type and elevation is done in order to find out if the model and the observation site had similar characteristics and if the station is not situated on a too steep terrain (decreasing the reliability in the comparison). Model and observed soil moisture variables (x) are normalized (xn) function of the minimum and maximum values of the time series (xmin
and xmax) (Brocca et al., 2013), according to the relationship shown in equation 4.6.
xn= x xmin
xmax xmin (4.6)
Even if the model has been run with four different soil layers, only two soil layers (SM1= 0.1 m and SM2=0.3 m soil depth) are compared, according to the available soil
depth observations. When the model soil layer and the measurements were at different depths, the observations has been interpolated at the same model depth using a linear interpolation. A spin-up time of two months is considered for the soil moisture results, according to the recommendations of the previous studies of Senatore et al. (2015). Figure 4.20 shows an example for the soil moisture comparison at Petrelle station. The reader is referred to the appendix for the full set of figures for all the stations in the catchment C.3.Soil moisture comparison exhibits a very good accordance in the time series, especially during the dry season in which the model manage to dry completely and with a very similar timing to the observations. Moreover the correlations in the seasonal peaks are quite well reproduced over all the year 2012. On the contrary, an anomalous behavior is observed during winter for all the model station sites in the most superficial soil layer (SM1). The model tends to dry suddenly, contrary to what observed. From a careful analysis it is evident that this kind of model behavior is observed every time that the temperature in the model goes under zero (Fig. 4.20, panel (c)). In this occasion, the model activates a frozen soil routine that, in my basin, seems to produce a non-observed drying of the most superficial soil layer.
Figure 4.20: Soil moisture dynamic for level SM1=0.1 m (panel (a)) and SM2=0.3 m (panel (b)) soil depths and rain and temperature variations (panel (c)) for the year 2012 at Petrelle station site.
Mean values, variance and correlation coefficients are reported in Tab. 4.5 for all the soil moisture station sites for the two soil depths. Mean values shows that, on average, there is not a clear overestimation or underestimation of soil moisture content of the model, compared with the observations. Variance varies in a range between 0.02 and
0.06 and shows a good accordance in terms of distribution in most of the stations, except San Benedetto, Pieve Santo Stefano and Foligno station sites for both soil moisture levels.
All the correlations coefficients have values in the range 0.63-0.95. The worst results are gained for the PgIng1 and PgIng2 stations sites and the best correlation are observed for Ficulle and Solomeo stations.
Table 4.5: Mean, variance and correlation coefficients for soil moisture comparison be- tween the best calibration run (WRF-Hydro) and observations (OBS) at all the stations sites inside the basin for the year 2012.
SM1=0.1 m
Mean Variance Correlation
Stations WRF-Hydro OBS WRF-Hydro OBS
T orre Olmo 0.49 0.25 0.06 0.06 0.69
Solomeo 0.45 0.39 0.06 0.08 0.90
S. Benedetto 0.47 0.57 0.05 0.10 0.91
P ieve S. Stef ano 0.52 0.65 0.03 0.12 0.82
P etrelle 0.52 0.36 0.04 0.06 0.84
M onterchi 0.49 0.52 0.05 0.08 0.89
F oligno 0.40 0.52 0.06 0.12 0.89
F iculle 0.39 0.56 0.06 0.06 0.95
Cerbara 0.54 0.31 0.06 0.08 0.75
P gIng1 0.23 0.52 0.03 0.03 0.66
P gIng2 0.23 0.39 0.03 0.07 0.69
SM2=0.3 m
Mean Variance Correlation
Stations WRF-Hydro OBS WRF-Hydro OBS
T orre Olmo 0.51 0.33 0.06 0.07 0.76
Solomeo 0.44 0.49 0.05 0.10 0.94
S. Benedetto 0.51 0.61 0.05 0.12 0.89
P ieve S. Stef ano 0.47 0.68 0.04 0.09 0.91
P etrelle 0.44 0.26 0.04 0.04 0.85
M onterchi 0.49 0.52 0.05 0.07 0.90
F oligno 0.42 0.52 0.06 0.12 0.87
F iculle 0.46 0.56 0.06 0.09 0.90
Cerbara 0.47 0.31 0.05 0.05 0.74
P gIng1 0.19 0.52 0.02 0.03 0.71
P gIng2 0.19 0.39 0.02 0.08 0.63
Except when frozen soil routine is activated, the run calibrated over the hydrograph
section, seems to reproduce a reliable soil moisture dynamic in all the observations sites at both soil depths.
4.4.6 Evaluation over 2013
My best calibration run (REFKDT= 0.3 and SATKDT= default + 0.9x10-6) has been evaluated over the following year, 2013. The hydrograph in Fig. 4.21 shows a good ability of the model in reproducing the main features of the seasonal water cycle, especially after the wet season and before the following dry season (from January to June) .
Figure 4.21: Comparison between the best calibration run with SATKDT=default + 0.9 x 10-6 and REFKDT=0.3 (blue line) and the streamflow observations (red line) at the Monte Molino river section for the period 2012 (calibration)- 2013 (validation) (m3/sec) The model manages to simulate the flood peak on the following year, even if overesti- mated (max modelled flow of 1359.8 m3/s, vs. 834.43 m3/s observed on the 2013 flooding event). The model produces a peak on 2013 that is even higher than the 2012 modelled flood peak. Nash Sutcliffe value of 0.75 and RMSE value of 51.81 (even lower than the calibration period) confirm the good performance of the model also for the validation period (Tab. 4.6). Even if the evaluation results are quite good, the model continue to overestimate low streamflow values in summer season (from the end of June to October) and to dry faster from the highest peaks. The possible effect of an higher saturation degree in the basin is reflected also in the validation period (2013) where it is evident that the model predicts higher flood peak than the observed ones and creating "fictitious floods" in the summer period. In this case the soil is simulated wetter than the actual one, creating more surface runoff and less contribution of groundwater (reflected in the
Table 4.6: Summary statistics for the best calibration run (2012), validation run (2013) and for the total analysis period (2012 and 2013)
Year REF KDT RET DEP RT F AC SAT KDT Q
Peak Q
Obs T peak T Obs
2012 0.3 default default
+0.9x10-6 1354.3 1337.9 13/11/12 02:00 13/11/12 15:00
2013 0.3 default default
+0.9x10-6 1359.8 834.4 12/11/13 08:00 12/11/12 15:00 2012-
2013
0.3 default default
+0.9x10-6 1359.8 1337.9 12/11/13 08:00 13/11/12 15:00
Y ear REF KDT RET DEP RT F AC SAT KDT RMSE RHO Nash-Sutcliffe
2012 0.3 default default+0.9x10-6 55.88 0.88 0.77
2013 0.3 default default+0.9x10-6 51.81 0.87 0.75
2012-2013 0.3 default default+0.9x10-6 53.66 0.88 0.77
different tails of recession limb). In order to investigate the possibility of an oversatura- tion of the model during the dry season, an soil moisture analysis analogous to section 4.4.5 is carried out for the year 2013.
In Fig. 4.22 is shown the soil moisture comparison at the Petrelle station site for hourly values. The full set of figures for the all the soil moisture stations is reported in the appendix C.4. The soil moisture station analyzed are the same analyzed in the calibration period and illustrated in Fig 4.4, except for Solomeo and Foligno stations that during the year 2013 were dismissed.
Figure 4.22: Hourly soil moisture dynamic for level SM1=0.1 m (panel (a)) and SM2=0.3 m (panel (b)) soil depths and rain and temperature variations (panel (c)) for the year 2013 at Petrelle station site.
Tab. 4.7 shows the correlation coefficient for the 2013 comparison.
Table 4.7: Mean, variance and correlation coefficients for soil moisture comparison be- tween the best calibration run (WRF-Hydro) and observations (OBS) at all the stations sites inside the basin for the validation year 2013 at hourly scale.
SM1=0.1 m
Mean Variance Correlation
Stations WRF-Hydro OBS WRF-Hydro OBS
T orre Olmo 0.59 0.67 0.04 0.11 0.91
S. Benedetto 0.57 0.55 0.06 0.11 0.92
P ieve S. Stef ano 0.55 0.56 0.02 0.11 0.80
P etrelle 0.51 0.42 0.03 0.07 0.86
M onterchi 0.54 0.52 0.05 0.07 0.86
F iculle 0.58 0.59 0.07 0.06 0.86
Cerbara 0.62 0.47 0.03 0.07 0.76
P gIng1 0.29 0.77 0.03 0.01 0.58
P gIng2 0.30 0.53 0.03 0.05 0.58
SM2=0.3 m
Mean Variance Correlation
Stations WRF-Hydro OBS WRF-Hydro OBS
T orre Olmo 0.59 0.53 0.06 0.06 0.84
S. Benedetto 0.56 0.63 0.07 0.14 0.93
P ieve S. Stef ano 0.46 0.60 0.04 0.06 0.87
P etrelle 0.46 0.34 0.06 0.05 0.95
M onterchi 0.57 0.56 0.07 0.05 0.95
F iculle 0.60 0.59 0.09 0.06 0.89
Cerbara 0.50 0.40 0.04 0.05 0.88
P gIng1 0.28 0.81 0.03 0.01 0.73
P gIng2 0.28 0.56 0.03 0.05 0.68
Correlation coefficients values are included in the range 0.58 to 0.94 (Tab. 4.7). The lowest values of correlation are scored for the station sites of PgIng1 and PgIng2, while the highest values are gained by the Petrelle station. Even if the WRF-Hydro stand- alone run is calibrated over 2012, the results in terms of soil moisture shows still a very good agreement with the observations also in the validation period. In 2013 comparison, similar results than the 2012 run in terms of soil moisture are highlighted. Also during the validation period the model doesn’t show a systematic over saturation of the terrain in the driest season, but at the contrary seems to dry with a similar tendency of the observations. Looking at the mean values of the soil moisture distributions, an overall improvement between the calibration and the validation run can be noticed, suggesting the fact that the soil moisture present a better dynamic after one year of simulation, after
the recession from the main flood peak of 2012. The correlation coefficient are high and the timing of the soil moisture variation of the model well correspond to the observed.
Similarly to 2012, the model has the same unexpected draining during winter season that depends on when the temperature goes under zero and the frozen soil routine of the model influence the most superficial soil layer.
Overall, the validation over the 2013 shows a good capacity of the model to reproduce the main features of the hydrological processes on the following year in terms of stream- flow and soil-moisture dynamic. The good results obtained proved that the calibrated parameter setting is able to reproduce not only the specific calibration year, but also to extend the experiment to other study cases over the area.
4.4.7 Preliminary conclusions over WRF-Hydro stand alone calibra- tion
The results obtained with the WRF-Hydro stand alone approach indicate that the model forced by the observation is able to satisfactorily reproduce the main flow peak and the timing of the hydrograph during the wet season (Fig. 4.16). Even if the model still overestimates low peak flows occurring in the driest season (from January to August), the overall skill of the best simulation calibration (REFKDT=0.3 and KSAT=+0.9x10-5) reaches a Nash Sutcliffe efficiency index of 0.77 and a total correlation of 0.88 (Tab.1). At the same way, the validation period the statistics results in a high Nash-Sutcliffe (0.75) and good correlations, with a resulting performance that is comparable to the calibration period.
A possible explanation to the flow overestimation of the low-discharge peaks may be explained by a model systematic over saturation of the terrain during dry season or by the fact that the model is calibrated over the one single year. The 2012 is in fact on one hand a very interesting year since both an extreme flooding event and an extreme drought are observed, but on the other hand the two extremes that occurs in that year may represent an obstacle to the calibration in properly reproduce the representativeness of average seasonal behavior of the basin. Finally, it has to be taken into account also the fact that low flow variations can be missed in the hydrograph due to the lack of hydrometer sensitivity or due to the presence of reservoirs upstreams that had to store water during the dry season, especially in one year in which an extreme drought is occurring (in my case the Montedoglio dam, total storage volume of 153x106 m3).
The analysis of the soil moisture highlights a good ability of the model to reproduce the main seasonal variation of the water soil content, even if an underestimation of the soil moisture occurs at the superficial soil layer when the frozen soil routine of the model is activated. In particular, both on the years 2012 and in 2013, the soil moisture dynamic does not show a systematic oversaturation of the model during summer. The WRF-Hydro stand alone runs are able to dry completely the water in the soil during the dry period. The soil moisture comparison helps to exclude the hypothesis of a lack of predictive ability of the model in reproducing soil moisture dynamic in summer months.
Therefore, the choice of the specific year 2012 for the calibration results to be the main reason of the difficulties of eliminating summer season streamflow overestimation.
The simulated soil conditions seems to produce more surface runoff and less contribution of groundwater (tails of recession limb). This is particularly evident looking at Fig.
4.21 for the period Jul-Oct (higher runoff) and Nov-Dec (lower sub-surface runoff) for both years 2012 and 2013. This shows that the WRF-Hydro is robust in modelling the hydrological processes (conserving the water balance), but a great care is needed for the parametrization choices. In this case a more extended calibration over multiple years could have corrected some model overestimations.
In general, when the final target of the study is to calibrate a model to optimize performance of low flow vs. high flow vs. average flow metrics, different optimization functions (e.g. log N-S vs. NS) and different parameters are used. In this experiment we a adopted a more general calibration approach that did not adapt a formal parameter sensitivity analysis for a specific optimization function, but explored single parameters sensitivity using manual calibration strategies (similar to Yucel et al. (2015)). Further improvements can be explored adopting an automated optimization analysis, instead than the manual calibration strategy used in this thesis.
It is possible to conclude that, overall, the calibration permits to the WRF-Hydro model to have a good predictive ability for the most representative processes of the water cycle over the area, considering multiple aspects of terrestrial hydrology processes in term of streamflow, fluxes and soil moisture variables. Since the main target of calibration is to prepare the WRF-Hydro setting in order to perform the fully coupled experiment, and not a specific hydrological calibration experiment over the area, the results are satisfactory enough to move to the next step.
4.5 WRF and WRF-Hydro fully coupled experiment
In the fully coupled experiment the output of the WRF and WRF/WRF-Hydro fully coupled runs (section 4.5.1) are analyzed in terms of precipitation (section 4.5.2), evap- otranspiration fluxes (section 4.5.4), soil moisture dynamic (section 4.5.3), runoff par- titioning (section 4.5.5) and streamflow (section 4.5.6) at different temporal scales. In particular the discussion is focused on understanding the eventual contribution of the WRF/WRF-Hydro hydrological routing and redistribution of water at the surface to precipitation predictability and, more in general, the water cycle processes (/section 4.5.7). The results are discussed in the following sections.
4.5.1 Model settings
The WRF and WRF/WRF Hydro fully coupled model settings are chosen accordingly to the previous findings of the WRF stand-alone experiment (section 4.3) and the WRF- Hydro stand-alone calibration (section 4.4) over the Tiber river basin. The domain of the simulations used for both the model runs is shown in figure 4.2 and discussed in detail in section 4.2.1.
The Kain-Frisch convective closure is used over the external domain at 12 km and convection is explicitly resolved in the d02 domain at 4 km. The Thompson microphysics
Figure 4.23: Averaged daily rainfall comparison over the Tiber river basin for year 2012.
scheme is chosen for the simulations, according to the discussion of section 4.3. The choice of the radiation schemes, vertical levels, PBL scheme are set accordingly to the specifications of section 4.3.
As introduced in section 2.3, the Noah-MP LSM model is used in the fully coupled experiment. LSM parameters are set accordingly to the results of the WRF-Hydro stand alone calibration experiment (section 4.4.7). At the same way, CORINE land cover dataset is used in the fully coupled experiment, as a result of the calibration phase (section 4.4.2).
The time step of the WRF model is 15 seconds for both WRF and WRF/WRF-Hydro runs on the meteorological grid corresponding to d01 domain at 12 km grid spacing; the time step on the high resolution terrain routing grid for the WRF-Hydro processes is 5 seconds, the same as specified in the calibration experiment.
Overland flow routing, channel routing and baseflow module are activated for the routing calculations.
4.5.2 Precipitation analysis
Figure 4.23 shows the average daily rainfall in the Tiber river basin for the year of the simulation. Rainfall amounts exhibit very small differences at the daily scale between WRF and WRF/WRF-Hydro simulations.
Highest differences are observed between each model simulation and the rainfall ob- servations than between the WRF and the WRF/WRF-Hydro fully coupled approach.
The differences in the comparison are more evident if we consider the daily accu- mulated rainfall values averaged over the domain. During spring and summer season
Figure 4.24: Averaged daily rainfall comparison over the Tiber river basin.
(from April to August) the WRF/WRF-Hydro cumulated rainfall results to be higher than the WRF stand alone (Fig. 4.24, panel (a)). Panel (b) of Figure 4.24, shows the daily rainfall differences (WRF/WRF-Hydro minus WRF daily rainfall) averaged at the catchment scale. Even if the differences are of the order of 3 mm/day, the graph high- light a tendency of the WRF-Hydro model to produce systematically more rainfall during spring and summer season (from April to mid June). In the following part of the year, where the contribution of the large scale is predominant, the differences became higher in magnitude but without a clear signal of one configuration that is producing higher or lower daily rainfall at the daily scale than the other.
Another important contribution to the discussion is given by the RMSE pixel-by-pixel differences at the daily scale over the Tiber river basin (Fig. 4.25). The RMSE in this case is not used as an evaluation method between a model run and the observation, but as an evaluator of the differences between WRF/WRF-Hydro (xcp) and WRF (xuncp) configurations (equation 4.7).
RM SE =
sPnpixels
n=1 (xcp xuncp)2
npixels (4.7)
Contrary to the previous analysis, in this case the rainfall fields are not averaged (and consequently smoothed) at the basin scale, but the contribution is calculated for every pixel inside the basin. Moreover, since the RMSE involves the absolute difference between the pixels, the differences are not compensating in sign. Fig. 4.25 highlights highest differences in the comparison for the end of July and during fall (September) and
Figure 4.25: pixel-by-pixel daily rainfall RMSE between WRF/WRF-Hydro and WRF over the Tiber river basin.
in the winter season from November to January, even if less intense. Especially in the last part of the year, higher discrepancies correspond to the most intense precipitation events. If we analyze the days with the highest daily rainfall differences in the analysis of figure 4.24 (panel (b)) and figure 4.25, we can isolate a set of five days that emerge both from the averaged analysis over the basin and the RMSE evaluation. In this way it is possible to look more in deep at the event scale, comparing the daily rainfall maps of the single days extrapolated from the seasonal runs with the observations.
Tab 4.8 summarize the days with highest discrepancies and figure 4.26 shows an example of the map comparison over one single day (the 14 September 2012). The full set of figures for the five days is reported in the appendix D.1.
Table 4.8: Summary of the days with highest differences between WRF/WRF-Hydro and WRF from the daily differences analysis and RMSE evaluation.
Days of maximum daily differences July 24th
September 3rd September 14th
October 13th N ovember 18th
Figure 4.26: Daily rainfall map comparison for the 14 September 2012 among WRF/WRF-Hydro (panel (a)), WRF (panel (b) and the gauge observations (panel (c)).
In general, the days with highest differences tends to produce similar patterns in coupled and stand alone configurations, with some changes in intensity and positioning.
The main physical characteristic of the precipitation processes remain the same. On 14 September (Fig. 4.26), the WRF/WRF-Hydro fully coupled (Fig. 4.26, panel (a)) seems to be more in accordance with the rain gauge interpolated observations, both in terms of rainfall quantity and localization. The two main rainfall peaks over the Adriatic coast are well reproduced, together with the very local rainfall spell located in the Irpinia region at the border with Puglia. The WRF stand alone model can capture the same rainfall characteristics as well, but producing an overestimation over the most intense precip- itation areas and an underestimation in the southern part of the Marche region, with maximum difference with the WRF/WRF-Hydro of 100 mm in the most intense rainfall core. Focusing on the Tiber river catchment area, the WRF stand alone (Fig. 4.26, panel (b)) tends to produce more rainfall in the east part of the basin with more intense and concentrated precipitation. The WRF/WRF-Hydro run, on the contrary, produces a widespread low intensity rainfall in the central part of the basin, more in line with the observations. In general, the differences between the model configurations (WRF/WRF- HYDRO and WRF stand alone) are lower than the differences of the model with the observations themselves (Fig. 4.26, panel (c)). The coupling seems not able to change significantly the main characteristics of the precipitation processes. This results are in line with the precipitation analysis of Senatore et al. (2015) that found similar results aver the Crati river Basin. The low feedbacks are mainly explained by the geographical location of the Italian peninsula (and in general of the Mediterranean region) in which the the weather systems are more influenced by large scale dynamics than by local pre- cipitation and fluxes. Since figure 4.25 and figure 4.26 shows appreciable differences in the spatial distribution of rainfall, while the accumulation line plot of figure 4.24 shows fairly minimal area integrated differences, it would be interesting to apply the MODE analysis in future applications to evaluate the skill of simulated precipitation features in these runs, as previously applied in the Pakistan flood 2010 experiment (see section 3.5.1 and section 3.6).
Finally, the precipitation analysis is analyzed also at the monthly and annual time scale. The monthly accumulated differences between the two model configurations and
Table 4.9: Monthly rainfall differences (mm) averaged over the Tiber river basin be- tween WRF/WRF-Hydro and observations (first column), WRF and observations (sec- ond columns), WRF/WRF-Hydro and WRF (third column).
Months Differences (WRF/WRF-Hydro-
Obs)
Differences
(WRF-Obs) Differences
(WRF/WRF-Hydro- WRF)
Jan 0.19 -1.09 1.28
F eb 47.33 44.68 2.65
M ar 7.86 7.81 0.05
Apr -13.18 -17.51 4.33
M ay -5.37 -8.64 3.27
Jun 2.06 2.21 -0.15
Jul 22.34 24.22 -1.88
Aug -18.85 -17.61 -1.24
Sep -17.33 -10.82 -6.51
Oct -54.33 -49.60 -4.73
N ov -114.17 -116.19 2.01
Dec -23.59 -23.30 -0.29
the observations are shown in Table 4.9.
Also in this case the discrepancies between the fully coupled and stand alone approach are lower than the differences between each single model run with the observations. In the comparison with the observations, the WRF/WRF-Hydro and WRF run tends to perform a strong overestimation of liquid precipitation in February 2012 (when an intense snowfall has characterized all the central part of Italy) and an overestimation during the month of November (when the main Tiber river flood occurred). The highest difference between WRF/WRF-Hydro fully coupled and WRF stand alone are observed in the month of September, where the fully coupled configuration produce 6 mm less than the WRF stand alone configuration over the average monthly rainfall. In the discussion of Table 4.9, it has to be taken into account that the results can be strongly influenced by differences in timing and localization of single rainfall days over the months, as considered in the previous discussion (Fig. 4.24 and Fig. 4.25).
Figure 4.27 shows the comparison of the pixel-by-pixel differences of the total accu- mulated rainfall over the full lenght of the simulation between WRF/WRF-Hydro and WRF run.
This evaluation permits to compare the annual rainfall and to identify if there are some preferred precipitation patterns in the two different model simulations. A tendency of WRF/WRF-Hydro to produce larger rainfall amount at yearly scale over the Adriatic coast, with differences peaking up to 200 mm. If we focus on the Tiber river basin, the WRF/WRF-Hydro tends to produce lower precipitation amounts over the upper part of the basin and higher precipitation values over the final part of the basin, close to the closing section in Ostia. The rainfall orographic pattern is evident, especially over the valley region of the Tiber river.
Figure 4.27: Annual accumulated rainfall differences map between WRF/WRF-Hydro and WRF for the year 2012 over the d02 domain.
If on one hand the yearly differences can be significant over some areas (10/15 % of the total average annual rainfall), the single daily rainfall differences are still modest.
Nevertheless higher differences at the daily/weekly scales are evaluated in terms of soil moisture dynamics, evapotranspiration and runoff. The impact of the coupling seems to be more significant in the framework of water resources management than at the single event scale.
4.5.3 Soil Moisture
Soil moisture dynamic is analyzed for the WRF and WRF/WRF-Hydro runs, compared with the observation measurements in the basin (Fig. 4.4). An additional term in the comparison is provided by the WRF-Hydro stand-alone calibration run described in section 4.4.5 (same settings as the WRF/WRF-Hydro run but with meteorological forcing given by the observations). The results of the comparison are displayed in figure 4.28 for the Petrelle station site, for the two soil layers SM1 and SM2 at 0.1m and 0.3 m of soil depth, respectively. The complete set of graphs for all the stations inside the catchment is provided in the appendix D.2.
Figure 4.28: Soil moisture comparison at the Petrelle station site for the different simula- tions for the SM1 and SM2 soil layers: WRF-Hydro calibration run (blue), WRF/WRF- Hydro (yellow), WRF (violet) and observed (red).
Also in this part of the experiment, the soil moisture time series have been normalized, according to method discussed in section 4.4.5, equation 4.6. In the same way, the first two months of the soil moisture dynamic are considered as spin-up (according to Senatore et al. (2015)).
The fully coupled WRF/WRF-Hydro run tends to better represent the drying period (from June to September), especially in the deeper soil layer SM2. The soil moisture evolution of the WRF/WRF-Hydro simulation seems to be more reactive than the WRF stand alone: when the model tendency is to increase the soil moisture after a drying period, the WRF/WRF-Hydro run tends to dry more and faster and to create a following wet peak that is higher than the stand alone configuration. In general, this simulation tends to be closer to the WRF-Hydro calibration stand alone run with perfect forcing.
It is also important to remember to the reader the fact that the WRF and WRF/WRF- Hydro forcing rainfall fields are coming from two different atmospheric model realizations, while the calibration run is forced by meteorological fields derived from the observations.
Since it is difficult to understand from these graphs (e.g. Fig 4.28) whether the variation of the soil moisture dynamic among the different simulations is originating from the con- tribution of the routing to the soil moisture distribution or by a difference in precipitation over the station site, an additional analysis is performed considering the different pre- cipitation contributions over the station sites. One time period in summer and one time period in winter are zoomed from the original time series of the soil moisture measure- ments sites and compared with the corresponding precipitation contributions from the
different forcings (WRF, WRF/WRF-Hydro and observed rainfall) in the nearest pixel.
In this way is it possible to understand if the soil moisture variations depend on the different precipitation forcing over the station site or on the improved routing dynamic that consider also lateral redistribution. Figure 4.29 shows the results of the analysis for the period July 15th to August 15th and figure 4.30 shows results from October 10th to December 1st for the Petrelle station site, level SM2. An intense summer rainfall spell and a more temporally distributed synoptic rain period are analyzed in the two time intervals, respectively. In both cases the WRF/WRF-hydro run exhibit a more reactive soil moisture dynamic. When the rain contributions are of equal intensity in the fully coupled and the stand alone configuration, the WRF/WRF-Hydro run tends to increase soil moisture faster and dry faster than the other configuration. This is especially evident in figure 4.29, when the WRF/WRF-Hydro run increase more and drain faster after the main precipitation event.
In addition to that, in figure 4.29, the rise of soil moisture on July 24th is lower than the observed one, although the simulated rainfall is much higher than the observed one.
The reason can be found in the fact the initial conditions of model are more wet than the observed ones, with a lower increase of saturation for the simulated soil moisture.
For the same period, the WRF stand alone doesn’t react to the rainfall at all, and SM continues to drop.
In addition to that, some characteristics of the soil moisture dynamic are closer to the observations in the fully coupled approach, thanks to the contribution of routing. An example of this behavior is particularly evident in figure 4.30 for the peak happening in the first days of November. Even if the observations and the WRF/WRF-Hydro have different values of soil saturation, the fully coupled run can reproduce well the shape of the double peak, more in line with the observations.
The mean values, variance and correlation coefficients associated to the different yearly simulations are shown in table 4.10, table 4.11 and table 4.12, respectively. Mean values of the different distributions does not give a clear signal of a systematic over- estimation/underestimation of the WRF and WRF/WRF-Hydro simulations with the observations. Variance in the distribution is generally higher for the observations than for the WRF and WRF-Hydro model simulations. In general, the difference between the WRF/WRF-Hydro and WRF are lower than with the observations. The correlation coefficients values are higher than 0.63 in all the stations, except than the Foligno sta- tion site in which the correlation values are really poor (also resulting uncorrelated in SM2). Since the results in the calibration part exhibit good correlation coefficients, a possible station malfunctioning in that year is excluded. This bad correlation is probably originated from a difference in term of precipitation of the two model runs over the area.
The WRF/WRF-Hydro fully coupled correlation coefficients results to be always higher than the WRF stand alone results, except than for three station over the total of eleven (Torre dell’ Olmo, Pieve Santo Stefano and PgIng2) in both soil layers SM1 and SM2. Overall, the WRF-Hydro improvements seems to be slightly better for the deeper soil layer SM2. The reason why in some cases the correlation coefficients of the calibration runs (forced by observations) are lower than the ones derived from the model
Figure 4.29: Soil moisture comparison at the Petrelle station site for summer period (from July 15th to August 15th) at SM2. Panel (a) shows the soil moisture content for the different simulations and as observed. Panel (b) shows the associated daily rainfall for the different simulations and raingauge observations.
Figure 4.30: Soil moisture comparison at the Petrelle station site for fall period (from October 10th to December 1st) at SM2. Panel (a) shows the soil moisture content for the different simulations and as observed. Panel (b) shows the associated daily rainfall for the different simulations and raingauge observations.
Table 4.10: Mean distribution values of modelled soil moisture of WRF/WRF-Hydro and WRF with the observations for the two soil moisture depths SM1 and SM2.
SM1=0.1 m
WRF/WRF-Hydro WRF WRF-Hydro calibr. Observations
T orre Olmo 0.43 0.52 0.49 0.25
Solomeo 0.47 0.45 0.45 0.39
S. Benedetto 0.49 0.45 0.47 0.57
P ieve S. Stef ano 0.48 0.47 0.52 0.65
P etrelle 0.48 0.46 0.52 0.36
M onterchi 0.48 0.45 0.49 0.52
F oligno 0.23 0.23 0.40 0.52
F iculle 0.42 0.47 0.39 0.56
Cerbara 0.48 0.46 0.54 0.31
P gIng1 0.42 0.49 0.23 0.52
P gIng2 0.42 0.49 0.23 0.39
SM2=0.3 m
WRF/WRF-Hydro WRF WRF-Hydro calibr. Observations
T orre Olmo 0.38 0.41 0.51 0.33
Solomeo 0.46 0.48 0.44 0.49
S. Benedetto 0.49 0.53 0.51 0.61
P ieve S. Stef ano 0.51 0.55 0.47 0.68
P etrelle 0.40 0.46 0.44 0.26
M onterchi 0.46 0.50 0.49 0.52
F oligno 0.43 0.38 0.42 0.52
F iculle 0.57 0.54 0.46 0.56
Cerbara 0.51 0.53 0.47 0.31
P gIng1 0.43 0.41 0.19 0.52
P gIng2 0.43 0.41 0.19 0.39
Table 4.11: Variance of modelled soil moisture distributions of WRF/WRF-Hydro, WRF and observations for the two soil moisture depths SM1 and SM2.
SM1=0.1 m
WRF/WRF-Hydro WRF WRF-Hydro calibr. Observations
T orre Olmo 0.06 0.06 0.06 0.06
Solomeo 0.06 0.07 0.06 0.08
S. Benedetto 0.05 0.07 0.05 0.10
P ieve S. Stef ano 0.06 0.07 0.03 0.12
P etrelle 0.03 0.04 0.04 0.06
M onterchi 0.05 0.07 0.05 0.08
F oligno 0.03 0.02 0.06 0.12
F iculle 0.08 0.07 0.06 0.06
Cerbara 0.06 0.07 0.06 0.08
P gIng1 0.07 0.06 0.03 0.03
P gIng2 0.07 0.06 0.03 0.07
SM2=0.3 m
WRF/WRF-Hydro WRF WRF-Hydro calibr. Observations
T orre Olmo 0.05 0.06 0.06 0.07
Solomeo 0.05 0.07 0.05 0.10
S. Benedetto 0.06 0.09 0.05 0.12
P ieve S. Stef ano 0.08 0.09 0.04 0.09
P etrelle 0.04 0.07 0.04 0.04
M onterchi 0.08 0.10 0.05 0.07
F oligno 0.02 0.03 0.06 0.12
F iculle 0.07 0.09 0.06 0.09
Cerbara 0.09 0.10 0.05 0.05
P gIng1 0.06 0.06 0.02 0.03
P gIng2 0.06 0.06 0.02 0.08
Table 4.12: Correlation coefficients of modelled soil moisture of WRF/WRF-Hydro and WRF with the observations for the two soil moisture depth SM1 and SM2.
SM1=0.1 m
WRF/WRF-Hydro WRF WRF-Hydro calibr.
T orre Olmo 0.67 0.68 0.69
Solomeo 0.88 0.82 0.90
S. Benedetto 0.83 0.83 0.91
P ieve S. Stef ano 0.89 0.92 0.82
P etrelle 0.77 0.72 0.84
M onterchi 0.83 0.83 0.89
F oligno 0.28 0.28 0.89
F iculle 0.87 0.84 0.95
Cerbara 0.78 0.78 0.75
P gIng1 0.80 0.76 0.66
P gIng2 0.80 0.81 0.69
SM2=0.3 m
WRF/WRF-Hydro WRF WRF-Hydro calibr.
T orre Olmo 0.71 0.76 0.76
Solomeo 0.88 0.87 0.94
S. Benedetto 0.79 0.70 0.89
P ieve S. Stef ano 0.84 0.88 0.91
P etrelle 0.78 0.71 0.85
M onterchi 0.82 0.75 0.90
F oligno -0.26 -0.42 0.87
F iculle 0.90 0.79 0.90
Cerbara 0.67 0.65 0.74
P gIng1 0.84 0.80 0.71
P gIng2 0.86 0.87 0.63
meteorological forcing can be found in some oscillations of some soil moisture stations that produce some mathematical artifices in the calculation of the correlation coefficients.
4.5.4 Evapotranspiration
The WRF/WRF-Hydro and WRF runs are compared in terms of evapotranspiration (Fig. 4.31 and Fig. 4.31). The evapotranspiration is obtained from each of the two model configurations as the sum of the ground surface evaporation rate, transpiration rate and evaporation of intercepted water for every hour. The quantity obtained is averaged over the Tiber river basin. The two run are not compared with the observations, because, unfortunately, there are not flux station measurements for that period in the Tiber river basin (as described in section 4.2.2).
Figure 4.31: Average evapotranspiration comparison over the Tiber river basin between WRF/WRF-Hydro (blue) and WRF (black) configurations.
As evident from figure 4.31 and figure 4.32, the WRF/WRF-Hydro run tends to evap- otranspirate more for all the period of the simulation. From the cumulated values of figure 4.32 (panel (a)), the WRF/WRF-Hydro fully coupled run provide evapotranspiration of the order of 15%-20% higher than the WRF stand alone.
During the first two months from the initialization (January and February) the stand alone and fully coupled runs are very close, and only later they start to detach with a major contribution of the WRF/WRF-Hydro evapotranspiration. This behavior is in- line with the spin-up time required by the soil moisture and the routing in general to
Figure 4.32: Average accumulated evapotranspiration comparison over the Tiber river basin between WRF/WRF-Hydro (blue) and WRF (black) configurations (panel (a)) and daily differences (panel (b)).
start the lateral routing redistribution. The highest differences are observed in spring and summer season (from March to September).
Since the evapotranspiration represent an important component of the water balance, in the following sections the other contributions to the hydrological cycle are analyzed in terms of runoff partitioning and flow rate response.
4.5.5 Runoff partitioning
The different contributions of the routing to the total runoff are evaluated in this section.
The main differences between the stand alone WRF and the WRF/WRF-Hydro fully coupled consists in the improved representation of the lateral routing and the in the possibility of the water to infiltrate and exfiltrate on its way to the channels. The runoff partitioning in WRF stand alone is calculated as a result of the 1-D Noah-MP LSM infiltration partitioning over every pixel of the domain, with water that is modelled not to move inside the domain. At the contrary, in the WRF/WRF-Hydro run the water route across the high-resolution terrain grid and contributions to channel network and baseflow.
For a typical 1-D NoahMP runs without the routing components (such as in regular WRF stand alone) the surface runoff (SFCRNFF) and underground runoff (UGDRNFF) terms are averaged over the river basin.
The WRF/WRF-Hydro total runoff is obtained from the streamflow at the basin outlet, representing the total runoff. The surface runoff and subsurface components are given by the channel inflow values (QSTRMVOLRT) from the routing grid and the UGDRNFF term from the land surface model grid averaged over the basin.
The comparison is done for the Tiber river basin with closing section at Monte Molino, the same section from which streamflow contributions have always been analyzed. The results of the comparison for the runoff partitioning are shown in figure 4.34, panel (a). In order to have a clearer understanding of the runoff partitioning comparison between the WRF/WRF-Hydro and the WRF, panels (b) and (c) of the same figure (Fig. 4.34) shows the performance of the other terms of the water balance in terms of evapotranspiration and precipitation averaged over the watershed at Monte Molino. Finally, panel (d) of figure 4.34 display the comparison in terms of soil moisture dynamic, to have a better correspondence between the runoff generation of panel (a) and the soil saturation and infiltration processes.
The total contributions to the runoff in the fully coupled configuration are lower at the beginning of the simulations and starts to be higher from June to the end of the year.
During all the summer and fall the total runoff contribution is higher (until November) and then rises again after the main flood event. The total volume of runoff for the WRF/WRF-Hydro and WRF is the same until November, but differently distributed in time. Then, the two runoff contributions starts to detach, with an higher WRF/WRF- Hydro contribution.
The underground runoff is higher in the fully coupled run. Consequently, the sur- face runoff contribution is lower for the WRF/WRF-Hydro run, if compared with the WRF stand alone simulation. The surface runoff lower contribution for the fully coupled
Figure 4.33: Different contributions in terms of water balance for the year 2012. WRF and WRF Hydro comparisons in terms of cumulated runoff (panel (a)), cumulated daily evapotranspiration (panel (b)), cumulated daily rainfall (panel (c)) and average hourly soil moisture inside the basin (panel (d)).
configuration is confirmed from the previous results. As a matter of fact, if the contri- butions in terms of precipitation of the two runs are similar (Fig. 4.34 panel (c) and section 4.5.2 discussion) and the evapotranspiration is higher in the WRF/WRF-Hydro fully-coupled configuration (Fig. 4.34 panel (b) and section 4.5.4), it is reasonable to expect a lower surface runoff term for the fully-coupled. The water that infiltrates into the soil is higher in the fully coupled approach (Fig. 4.34 panel (d) and section 4.5.3 discussion), consequently there is an higher contribution in terms of subsurface runoff and more underground runoff, compared to the WRF stand-alone simulation.
4.5.6 Streamflow evaluation
Since one of the main peculiarities of the WRF/WRF-Hydro fully coupled approach is the possibility to obtain in a single simulation results that goes from the atmosphere to the terrestrial hydrology and flow values in the channel network, I conclude the experiment looking at the streamflow results.
The WRF/WRF-Hydro results in terms of streamflow are compared with the observa- tion at Monte Molino river section and with the streamflow obtained using WRF-Hydro stand alone forced by the WRF stand alone run. In order to produce this last run, the WRF-Hydro model (with the same calibrated settings of the other WRF-Hydro configu- rations) is forced by the WRF run. The results shown in figure 4.34 are reasonably good if we consider that the meteorological forcing is derived from the WRF model runs and not from observations. The Nash Sutcliffe value of the simulation is 0.34, in line with
Figure 4.34: Hydrograph comparison among WRF, WRF/WRF-Hydro and the observa- tions at the Monte Molino closing section, for the year 2012.
the same values of Nash Sutcliffe obtained in the previous studies( Givati et al. (2016), Senatore et al. (2015) and Arnault et al. (2016)).
The statistics are summarized in Table 4.13. The WRF/WRF-Hydro run outperform the WRF stand alone run in all the evaluators (RMSE, correlation coefficient and Nash- Sutcliffe index).
The higher differences between the two runs are observed for low flow discharge values, where the run forced by WRF stand alone tends to produce higher peaks. This behavior may be explained by the different frequency in the forcing. The WRF rainfall is given every hour by the LSM, while the fully coupled approach calls the LSM more frequently (time step of 15 seconds). For this reason, when an isolated rainfall spell falls in the basin, the WRF stand alone forced run has less time for the routing to operate, Table 4.13: Streamflow evaluation for the WRF/WRF-Hydro and WRF in terms of maximum peak (m3/s), time of the peak, RMSE (m3/s), RHO, Nash-Sutcliffe
Run Q
Peak Q
Obs T peak T Obs RMSE RHO Nash-
Sut.
WRF/WRF-Hydro 673.82 1337.9 12/11/12 20:00 13/11/12 05:00 93.72 0.74 0.34 WRF 664.06 1337.9 12/11/12 20:00 13/11/12 05:00 93.72 0.70 0.33
