Can we replace observed forcing with weather generator in land surface modeling? Insights from long-term simulations at two contrasting boreal sites

Abstract

This study evaluates the simulation of water balance components at half-hourly time steps from the Canadian Land Surface Scheme (CLASS) when driven by a 500-year stochastic meteorological data set produced by the Advanced WEather GENerator (AWE-GEN) at two boreal sites with contrasting water availability. The CLASS was driven by ERA5 reanalysis data (CLASS-CTL) over 39 years and its output was used as a surrogate for land surface observations. At both sites, the mean monthly and annual values of all meteorological variables used to drive CLASS, including precipitation, are well captured by AWE-GEN, but their variability is, sometimes, biased. In general, CLASS driven by stochastic data (CLASS-WG) tends to produce higher evapotranspiration compared to values simulated by CLASS-CTL, especially during spring and summer at the wet site. The interannual evapotranspiration-precipitation and runoff-precipitation relationships derived from CLASS-WG and those derived from CLASS-CTL were very similar to each other at the dry site; they both indicate that evapotranspiration and runoff are limited by water availability. At the wet site, however, CLASS-WG only captured well the interannual runoff-precipitation relationship. The sensitivity analysis shows that CLASS water fluxes are particularly affected by the replacement of physically consistent input time series of incoming short-wave radiation, precipitation, temperature, and specific humidity. In conclusion, the results show that even though a weather generator can produce coherent climate time series, the use of this synthetic data as meteorological forcing in a physically based land surface model does not necessarily reproduce the complex surface processes, such as the surface water fluxes. More studies are encouraged to further analyze the constraints of this framework.

Appendices

Appendix A. Local validation of ERA5 reanalysis data

Figures 10 shows comparisons between the daily mean atmospheric variables derived from the ERA5 reanalysis and observations at site SK-OAS from 1997 to 2010.

Fig. 10

Comparisons between in situ observations and ERA5 reanalysis data of (a) daily incoming short-wave radiation, (b) daily incoming long-wave radiation, (c) daily mean near-surface air temperature, (d) daily mean vapor pressure deficit, (e) daily mean wind speed, (f) daily mean surface air pressure, and (g) daily total precipitation at the SK-OAS site. The color bar indicates the density distribution of each scatterplot. Darker blue means a dense distribution, and the white color means sparse distribution. The statistical scores, including adjusted-R square, root-mean-square error, and relative bias, are also presented

Figures 11 displays comparisons between the daily mean atmospheric variables derived from the ERA5 reanalysis and observations at site QC-Juv from 2016 to 2018.

Fig. 11

Comparisons between in situ observations and ERA5 reanalysis data of (a) daily incoming short-wave radiation, (b) daily incoming long-wave radiation, (c) daily mean near-surface air temperature, (d) daily mean vapor pressure deficit, (e) daily mean wind speed, (f) daily mean surface air pressure, and (g) daily total precipitation at the SK-OAS site. The color bar indicates the density distribution of each scatterplot. Darker blue means a dense distribution, and the white color means sparse distribution. The statistical scores, including adjusted-R square, root-mean-square error, and relative bias, are also presented

Table 2 provides a summary of the statistical metrics associated with the comparison between ERA5 data and local observations at sub-daily scale at site SK-OAS.

Table 2 Statistical summary of comparison, at sub-daily scale, between observational and ERA5 reanalysis data at SK-OAS from 1997 to 2010. The atmospheric variables are incoming short-wave radiation (SW_in), incoming long-wave radiation (LW_in), air temperature (T_a), vapor pressure deficit (VPD), wind speed (Ws), surface pressure (Pre), and total precipitation (P). The statistical metrics are the mean values, root-mean-square error (RMSE), adjusted-R square (R²), and relative bias (pBias)

Table 3 serves the same purpose as Table 2 but applies to the QC-Juv site.

Table 3 Statistical summary of comparison, at sub-daily scale, between observational and ERA5 reanalysis data at QC-Juv from 2016 to 2018. The atmospheric variables are incoming short-wave radiation (SW_in), incoming long-wave radiation (LW_in), air temperature (T_a), vapor pressure deficit (VPD), wind speed (Ws), surface pressure (Pre), and total precipitation (P). The statistical metrics are the mean values, root-mean-square error (RMSE), adjusted-R square (R²), and relative bias (pBias)

Appendix B. Air specific humidity formulation

Specific humidity (q in g g^–1) can be calculated through its relationship with (unsaturated/saturated) vapor pressure and atmospheric surface pressure (P in Pa) (see Stull 2015):

$$ q=\frac{\varepsilon \cdotp e}{P-e\cdotp \left(1-\varepsilon \right)}, $$

(Eq. A1)

which e (Pa) is the actual (unsaturated) or saturated vapor pressure, and ε is equal to 0.622. We may assume the unsaturated specific humidity when air temperature (T_a in °C) is higher than the dewpoint temperature (T_d in °C). T_d is calculated as T_d = 273.156 + [1/T₀ − R_v/L_v ·  ln (e_a/e₀)]⁻¹, with T₀ = 273.156°C, e₀= 611 Pa, and R_v/L_v = 1.844 x 10^–4 K^–1. For a saturated atmospheric surface (T_a = T_d), we use saturated vapor pressure (e_s in Pa). e_s can be calculated as a function of T_a, such as e_s = e₀ exp [17.27T_a/(237.3 + T_a)].

Appendix C. Incoming long-wave radiation formulation

Surface incoming long-wave radiation (LW_in) is typically estimated by first determining the incoming long-wave radiation for clear-sky conditions (LW_in,c) and then correcting for cloud fraction (Wang and Liang 2009). LW_in,c can be calculated as

$$ {LW}_{in,c}={\varepsilon}_a\sigma {T_a}^4, $$

(Eq. B1)

where ε_a is atmospheric emissivity, σ is the Stefan-Boltzman constant (= 5.67 x 10^-8 W m^–2 K^–4), and T_a is the air temperature in K. According to Brutsaert (1975), ε_a can be empirically obtained as

$$ {\varepsilon}_a=a{\left(\raisebox{1ex}{${e}_a$}\!\left/ \!\raisebox{-1ex}{${T}_a$}\right.\right)}^b, $$

(Eq. B2)

where a and b are 1.24 and 1/7, respectively (Brutsaert 1975). However, these coefficients can be easily calibrated for a particular site. Once the LW_{in, c} values are obtained, the LW_in can be finally estimated as

$$ {LW}_{in}={LW}_{in,c}\left(1+c{N}^d\right), $$

(Eq. B3)

where N is the cloud cover fraction, ranging from 0 to 1, and c and d are empirical coefficients.

In this study, we calibrated the empirical coefficients a, b, c, and d for the study site using the LW_in, T_a, e_a from the ERA5 data set. Table 4 presents the values of the estimated coefficients a, b, c, and d for the sites SK-OAS and QC-Juv.

Table 4 Calibrated coefficient values, following Brutsaert (1975) formulation, for estimating the incoming long-wave radiation at the SK-OAS and QC-Juv

Figure 12 shows the estimated LW_in,c and LW_in, calibrated following Brutsaert's (1975) formulation for the sites SK-OAS and QC-Juv. Both empirically estimated LW_in,c and LW_in are highly correlated (Pearson correlation higher than 0.9) to incoming long-wave radiation from ERA5 data, with relative bias less than 1 %.

Fig. 12

Comparing CTL incoming long-wave radiation with empirically estimated values at (a) SK-OAS site for clear-sky condition only, (b) QC-Juv site for clear-sky conditions only, (c) SK-OAS site for all-sky condition, and (d) QC-Juv site for all-sky conditions. Cor, RMSE, and Bias are the statistical scores Pearson correlation, root mean square error (W m^-2), and relative bias (%), respectively. The color bars indicate the density distribution of each scatter plot. The red dashed line is 1:1 perfect fit

Finally, the LW_in empirical model with calibrated coefficients was used to estimate the variable using stochastic data generated by the AWE-WGEN model. Figure 13 compares the estimated LW_in from WG with CTL data. Overall, the density distribution of hourly values at SK-OAS and QC-Juv sites shows that both data sets, from both sites, are very well comparable to each other, with minimal differences in their mean (less than 2.5 W m^–2) and standard deviation (less than 8.0 W m^–2) values (Fig. 11a and 11b). The mean intra-annual variation of estimated LW_in from WG data is also very similar to LW_in from CTL (Fig. 11c and 11d), with more significant differences, around 13 W m^–2, in January and February at the sites SK-OAS and in July at QC-Juv, respectively.

Fig. 13

The probability distribution function of incoming long-wave radiation (LW_in, at hourly time-step) from WG and CTL dataset at (a) SK-OAS and (b) QC-Juv, and the mean intraannual variation of LW_in from WG and CTL data at (c) SK-OAS and (d) QC-Juv. The LW_in values from WG and CTL are given in red and black lines, respectively. E and σ, in (a) and (b), are the mean and standard deviation values, respectively. Vertical bars in (c) and (d) represent monthly standard deviation values

Appendix D. Annual cycles of sub-daily biases in stochastic meteorological variables

Figure 14 shows annual cycles of sub-daily (daytime and nighttime) normalized biases in stochastic meteorological variables produced at sites SK-OAS and QC-Juv. The biases are the differences between the monthly mean values of the stochastic (WG) and the control (CTL) data. The monthly differences are after normalized by the monthly mean value of the CTL variable.

Fig. 14

Annual cycles of sub-daily normalized bias of precipitation at (a) SK-OAS and (b) QC-Juv, incoming short-wave radiation at (c) SK-OAS and (d) QC-Juv, air temperature at (e) SK-OAS and (f) QC-Juv, vapor pressure deficit at (g) SK-OAS and (h) QC-Juv, and air specific humidity at (i) SK-OAS and (j) QC-Juv. Orange and blue lines represent daytime and nighttime values. The normalized bias is calculated as $\left(\overline{Var_{WG}}-\overline{Var_{CTL}}\right)/\overline{Var_{CTL}}$, which $\overline{Var_{WG}}$ is the monthly mean value of the stochastic variable (WG data) and $\overline{Var_{CTL}}$is the monthly mean value of the control variable (CTL data)

Appendix E. Simulation of annual precipitation partitioning into evapotranspiration and runoff

Figure 15 shows the simulated annual precipitation partitioning into evapotranspiration and runoff from CLASS driven by the stochastic and reference data at SK-OAS and QC-Juv.

Fig. 15

Pie charts of the annual precipitation partitioning into evapotranspiration (in green) and runoff (in grey) simulated by CLASS-WG at (a) SK-OAS and (b) QC-Juv, and by CLASS-CTL at (c) SK-OAS and (d) QC-Juv

Appendix F. Simulation of the evapotranspiration terms at the humid boreal site (QC-Juv)

Figure 16 presents scatter plots of the simulated interannual relationship of precipitation versus canopy evapotranspiration and precipitation versus ground evaporation and sublimation at QC-Juv. It must be highlighted that canopy evapotranspiration term refers to water loss from the canopy to the atmosphere by transpiration and by evaporation of intercepted precipitation. The ground evaporation and sublimation term refer to water loss from the bare and snow-covered ground to the atmosphere.

Fig. 16

(a) Scatter plot of annual cumulative precipitation (P) versus simulated canopy evapotranspiration at QC-Juv. (b) scatter plot of annual cumulative P versus simulated ground evaporation and sublimation at QC-Juv. Red and black dots and linear regression lines represent the variables from WG and CTL data, respectively

Appendix G. Meteorological forcing variables from stochastic and reference data set: incoming short-wave radiation, precipitation, air temperature, and specific humidity

Figure 17 presents the mean intra-annual variation of the precipitation, air temperature, and specific humidity used to drive CTL, CTL-P, and CTL-T_aq simulations.

Fig. 17

Mean intraannual variation of daily incoming short-wave radiation (SW_in) at (a) SK-OAS and (b) QC-Juv, daily precipitation (P) at (c) SK-OAS and (d) QC-Juv, daily mean air temperature (T_a) at (e) SK-OAS and (f) QC-Juv, and daily mean air specific humidity (q) at (g) SK-OAS and (h) QC-Juv. Boxplots represent the distribution of the daily values of SW_in, P, T_a, and q at each month (from January to December), and the solid curves are the monthly mean values. The variables displayed in black, light green, dark blue, and dark green colors were used to drive the CTL, CTL-SW_in, CTL-P, and CTL-T_aq simulations