Abstract

Using the Connecticut River basin as an example, this study assesses the extent to which remote sensing data can help improve hydrological modeling and how it may influence projected future hydrological trends. The dynamic leaf area index (LAI) derived from satellite remote sensing was incorporated into the Variable Infiltration Capacity model (VIC) to enable an interannually varying seasonal cycle of vegetation (VICVEG); the evapotranspiration (ET) data based on remote sensing were combined with ET from a default VIC simulation to develop a simple bias-correction algorithm, and the simulation was then repeated with the bias-corrected ET replacing the simulated ET in the model (VICET). VICET performs significantly better in simulating the temporal variability of river discharge at daily, biweekly, monthly, and seasonal time scales, while VICVEG better captures the interannual variability of discharge, particularly in the winter and spring, and shows slight improvements to soil moisture estimates. The methodology of incorporating ET data into VIC as a bias-correction tool also influences the modeled future hydrological trends. Compared to the default VIC, VICET portrays a future characterized by greater drought risk and a stronger decreasing trend of minimum river flows. Integrating remote sensing data with hydrological modeling helps characterize the range of model-related uncertainties and more accurately reconstruct historic river flow estimates, leading to a better understanding and prediction of hydrological response to future climate changes.

1. Introduction

Understanding and quantifying how climate variability and changes may influence water resources and the terrestrial hydrological cycle is of critical importance for socioeconomic development and remains a major challenge facing the field of hydrology. In many regions, studies of past hydrological variability are hampered by the lack of reliable river flow data. These include regions where no river gauge is maintained and regions where water diversion or hydraulic structure has altered the natural flow. Hydrological models provide an important tool for reconstructing and understanding past hydrological variability and for quantifying future hydrological changes.

Evapotranspiration (ET) is an important component of the terrestrial hydrological cycle. As an important pathway for land–atmosphere interactions, ET influences the surface water, energy, and carbon fluxes (Xia et al. 2015). However, ET is a particularly difficult process to accurately measure or simulate. In the study of Kite and Droogers (2000) using eight different methods of estimating actual evaporation and transpiration, including field data, hydrological models, and the early generation of satellite data, a wide range of ET estimates emerged from the various methods, and no method is evidently better than others. Another study (Weiß and Menzel 2008) compared four different equations for potential ET to assess their impact on streamflow simulation and found that results based on each equation differ substantially from one another. Together with many others on this topic (McKenney and Rosenberg 1993; Eitzinger et al. 2002; Lu et al. 2005; Bormann 2011; Trambauer et al. 2014), these studies highlighted the need to explore innovative approaches to improving the accuracy of ET estimation in hydrological models.

In a recent study on hydrological modeling for the Connecticut River basin (CRB) using the Variable Infiltration Capacity model (VIC; Parr and Wang 2014), modeled ET was compared against the ET data derived from Advanced Very High Resolution Radiometer (AVHRR) and the International Satellite Land Surface Climatology Project, Initiative II (ISLSCP-II; Fisher et al. 2008), and a significant underestimation of ET (especially in summer) was identified. The summer [June–August (JJA)] underestimation of ET by VIC identified by Parr and Wang (2014) ranged from 20% to 30% as compared to Fisher et al. (2008). It is not clear whether the underestimation of ET by VIC results from the estimation of potential evaporation by the Penman–Monteith equation used in the model (which requires several parameters difficult to characterize) or because of limiting soil moisture storage. A general underestimation of ET (and therefore overestimation of runoff ratio) was also documented in several other studies using VIC (e.g., Xia et al. 2015; Vano et al. 2012; Sheffield et al. 2012). Recently, Xia et al. (2015), on the basis of output from the North American Land Data Assimilation System, phase 2 (NLDAS-2), compared the output of four land surface models including VIC to eddy covariance data at 46 AmeriFlux tower sites and found that VIC had the lowest ET of all the models and was less than observation for most of the year, but particularly in the summer.

Given the model bias in simulating ET in the past, future hydrological predictions (e.g., Parr et al. 2015, manuscript submitted to Global Planet. Change) may also be subject to similar model biases, leading to substantial uncertainties in the model-generated future hydrological trend. In particular, a major challenge and uncertainty in modeling streamflow is related to the difficulty in realistically modeling the ET process. Although ET is biased low, VIC streamflow has been shown to be more accurate than the Noah or Mosaic models (Xia et al. 2012a) using the same NLDAS-2 forcing data. In this study, we integrate ET remote sensing data with hydrological modeling using VIC and examine how more accurate ET in the model may influence model performance in simulating other hydrological processes.

Vegetation is an important land surface state variable that profoundly affects both the long-term and seasonal dynamics of all components of the water cycle (Wattenbach et al. 2012). There are a number of vegetation parameters within VIC, including roughness length, displacement height, architectural resistance, minimum stomata resistance, and rooting depth. However, leaf area index (LAI) has the greatest influence on hydrologic simulations in VIC (Ford and Quiring 2013). In this study, we replace the model’s typical system of using a climatological or static LAI with an interannually varying dynamic LAI in order to improve the model simulation of interannual variability of hydrological processes.

A large number of past studies have applied remote sensing data to hydrological models. Dynamic vegetation parameters have previously been employed in land surface models (LSMs), typically in the form of LAI acting as a model input parameter for purposes such as investigating the impacts on long-term and seasonal dynamics of the water cycle (Wattenbach et al. 2012), on the predictability of ET patterns (Tang et al. 2012), on potential improvements of streamflow (Zhang et al. 2011; Zhou et al. 2013), and on soil moisture (Ford and Quiring 2013). Soil moisture has been employed in hydrological models as a forcing to improve estimates of actual ET (e.g., Ottlé and Vidal-Madjar 1994) and to improve runoff estimates and regional model calibration (e.g., Zhang et al. 2011). Remote sensing–based ET data have been used in past hydrological modeling, but primarily for model calibration purpose in identifying the model parameters that best reproduce the ET data (e.g., Immerzeel and Bold 2008).

Here, we propose a new approach to improve hydrological modeling and prediction using ET remote sensing data, in which the data are used to support a bias-correction procedure that can be applied to the model for any time period. Using the Connecticut River basin as an example, this study assesses the extent to which remote sensing data can help improve hydrological modeling and how it may influence future hydrological trends. In the following, section 2 provides a description of the datasets utilized and our experimental design. Section 3 describes the results, including both model improvements and the impact on hydrological trends with a focus on extremes such as flood and drought risks. Section 4 provides a summary considering model-related uncertainty and the sensitivity of the model results to the choice of the ET dataset, as well as a discussion based upon the results found.

2. Data and methodology

a. Datasets

The model used is VIC, version 4.1.2.g (Liang et al. 1994). In earlier studies, we have applied VIC to the Connecticut River basin for past simulation and future projections. The simulation for the past was driven by NLDAS-2 (Xia et al. 2012b,c) forcing data for the period 1980–2011 (Parr and Wang 2014), and future projections for the period 2046–65 were driven by downscaled and bias-corrected output from three North American Regional Climate Change Assessment Program (NARCCAP) models. The NARCCAP models (Mearns et al. 2012) from which the three future forcings were derived include the Regional Climate Model, version 3 (RegCM3; Pal et al. 2007), driven with lateral boundary conditions (LBCs) from the Third Generation Canadian Coupled Global Climate Model 3 (CGCM3; RegCM–CGCM); RegCM3 driven with LBCs from the Geophysical Fluid Dynamics Laboratory Climate Model, version 2.1 (GFDL CM2.1; RegCM–GFDL); and the Canadian Regional Climate Model (CRCM; Caya and Laprise 1999) driven with LBCs from CGCM3 (CRCM–CGCM). The future daily meteorological forcing from each RCM–GCM combination was first downscaled to ⅛° resolution and bias corrected using the methodology of Ahmed et al. (2013). Based on quantile mapping between the cumulative distribution function (CDF) of a modeled present-day climate and that of observations, the downscaling and bias-correction algorithm of Ahmed et al. (2013) was developed based on the bias correction and downscaling method in Wood et al. (2002, 2004), but with modifications for application to daily modeled output and modifications to the procedure to take full advantage of the fine-resolution observational data. While the meteorological forcing is at ⅛° resolution, the model simulations are run at a 0.025° resolution in order to incorporate fine-resolution topographical data and land-cover data. Therefore, all 25 grid cells within each ⅛° box share the same meteorological forcing. The surface water (and energy) balances in VIC are solved at each 0.025° grid cell.

The land-cover data in this study were taken from the University of Maryland’s Global Land Cover Facility; DEM data were taken from the USGS HYDRO1k project (USGS 2011); and soil data including soil type and texture, porosity, and bulk density were taken from the National Oceanic and Atmospheric Administration/National Geophysical Data Center (NOAA/NGDC; Reynolds et al. 2000). The vegetation data used are the LAI data derived from Moderate Resolution Imaging Spectroradiometer (MODIS) sensors on board the polar-orbiting Terra and Aqua satellites. The LAI data have an 8-day temporal resolution from which a monthly average is derived and a 1-km spatial resolution matching our land-cover dataset. The ET data, spanning 1986–95, were from Fisher et al. (2008) with monthly and 0.5° spatial resolutions. As ET flux cannot be remotely sensed, it is estimated according to a surface radiation budget algorithm with net radiation, air temperature, and water vapor pressure taken from the ISLSCP-II and visible spectrum reflectance and near-infrared spectrum reflectance gathered with AVHRR. This ET product has been validated over 36 FLUXNET towers sites, with a coefficient of determination R2 value of 0.90 and a 7% bias (Fisher et al. 2008, 2009). As eddy flux sites generally achieve about 90% closure for an energy balance with 10% of the variation unexplained, for a product like this we would not expect any R2 value much greater than 0.90 (Fisher et al.2008). Vinukollu et al. (2011) compared three remote sensing ET products [including the Fisher et al. (2008) product used here] and VIC and ERA-Interim estimates. It was found that the product used in this study had the lowest bias of annual ET over 26 basins and the highest correlations of monthly mean latent heat flux compared with tower observations.

For discharge simulation and analysis, USGS station 01184000 (USGS 2012) at Thompsonville, Connecticut, is used, which had been identified as a reliable station in previous research (Bjerklie et al. 2011; Marshall and Randhir 2008). Also used is USGS station 01144500 (USGS 2012), located north and upstream at West Lebanon, New Hampshire. Remote sensing data from the European Space Agency’s (ESA) Essential Climate Variable (ECV) surface soil moisture product are also utilized to compare potential model improvements. This product (Naeimi et al. 2009; Dorigo et al. 2010; Liu et al. 2012) is a merged dataset from both active and passive microwave measurements and has a spatial resolution of 0.25°.

b. Experimental design

Several different types of simulations are conducted. One uses the default VIC calibrated for the CRB (Parr and Wang 2014), referred to as “VIC.” VIC was calibrated using the USGS river discharge observations at Thompsonville for the first 10 years of each simulation (1950–60 and 1980–90, respectively). The suggested calibration parameters such as the variable infiltration parameter (b_infilt), the soil parameter controlling when nonlinear base flow occurs (Ds), the fraction of maximum soil moisture where nonlinear base flow occurs (Ws), maximum baseflow velocity (Dsmax), and soil layer depth were adjusted such that discharge best matched USGS observations in terms of both magnitude and variation on a variety of different time scales (Parr and Wang 2014). The model was then verified using discharge data at both the Thompsonville and West Lebanon stations for the rest of the simulation period as well as remotely sensed ET and soil moisture. A more descriptive explanation of the default VIC and its calibration and validation process for the Connecticut River basin can be found in Parr and Wang (2014).

VIC requires LAI data as model input; however, it uses the climatological LAI with a seasonal cycle and does not include interannual variation. This type of LAI condition is referred to hereafter as “static vegetation.” In a second type of simulation (VICVEG), for a more realistic representation of vegetation, the MODIS LAI data are used as model input. The MODIS LAI includes not only seasonal but also interannual variations of vegetation captured by the remote sensing data. This type of vegetation condition is referred to hereafter as “dynamic vegetation.” In simulations with dynamic vegetation, the model was adjusted to read in different seasonal LAI datasets for each year of a simulation (Fig. 1), although the fraction of various land covers does not change. Within the model, LAI quantifies canopy cover and influences the maximum amount of water intercepted by the canopy, canopy resistance, and root water uptake, thus affecting evapotranspiration rates (Liang et al. 1994). The use of the MODIS-derived LAI can therefore potentially improve the simulated seasonal and interannual variability of the surface water budget. It should be noted that since the climatological LAI is from a different source, the evaluation of performance and potential improvements will focus on temporal variations rather than the magnitude, particularly at the interannual time scale.

Fig. 1.

Seasonal comparison between the dynamic LAI used in VICVEG and the static LAI used in the default VIC.

Fig. 1.

Seasonal comparison between the dynamic LAI used in VICVEG and the static LAI used in the default VIC.

In another type of simulation, the model-simulated ET components were overwritten to correct the ET bias identified based on the comparison of ET from a default model simulation with ET data derived from remote sensing. This procedure therefore involves a pair of simulations. The default VIC calculates the potential ET based on atmospheric forcing using the Penmen–Monteith equation and then derives the actual ET based on water availability. The model calculates each of the five components of ET (canopy evaporation, transpiration, bare ground evaporation, canopy sublimation, and surface sublimation) separately, at daily or hourly time steps, while the remotely sensed ET data are monthly. In this study, the monthly average of modeled ET from a default run is compared with the “observational” ET (from remote sensing data) during 1986–95 to calibrate a ratio between the two, and this was done for each grid cell for each of the 12 months. The resulting ratio, which varies spatially and seasonally, is then used to scale the model-simulated five ET components (at the model native resolutions both spatially and temporally) for any period of interest assuming stationarity of the model–observation relationship. As the remote sensing–based ET data cannot distinguish the five different components of ET, all five are scaled equally, and this scaling ratio is specific to a grid cell and a particular month. The resulting data for the five ET components (referred to as the corrected ET components) are then used to overwrite the model-simulated values in a rerun of the simulation, and this type of simulation is referred to as “VICET.” As such, in the VICET simulation, the use of ET data prevents the inaccuracy of ET parameterization from propagating to other hydrological variables (including, for example, runoff and soil moisture), thus correcting the bias of the model hydrological processes caused by inaccurate ET parameterization. Figure 2 provides a schematic diagram describing the procedure of implementing the ET bias-correction algorithm into hydrological modeling.

Fig. 2.

A flowchart of the calibration and implementation of the ET bias-correction process utilized. Variable fet represents the month- and gridcell-specific relationship between model and observation, ETobs represents the remote sensing ET for the time period with remote sensing data available, ETmod represents the total model-simulated ET for the time period with remote sensing data available, and ETi represents each of the five ET components simulated in each time step by the default model over the application period of interest.

Fig. 2.

A flowchart of the calibration and implementation of the ET bias-correction process utilized. Variable fet represents the month- and gridcell-specific relationship between model and observation, ETobs represents the remote sensing ET for the time period with remote sensing data available, ETmod represents the total model-simulated ET for the time period with remote sensing data available, and ETi represents each of the five ET components simulated in each time step by the default model over the application period of interest.

Note that overwriting a model variable with a prescribed value often causes violation of mass and/or energy conservation. When a violation does occur, depending on specific processes related to such violation, a corresponding procedure will be triggered in the model to adjust the values of related variables (including the overwritten variable itself) to conserve mass and energy, thus maintaining physical realism. For example, if in a given time step the prescribed canopy evaporation amount is greater than the current storage of the canopy, it will be reduced back to the value of the canopy storage. Likewise, ground evaporation cannot exceed what is currently stored at the surface, and transpiration will not deplete soil moisture past the wilting point. As VIC derives its estimates of soil moisture and runoff after ET estimates, mass balance is compensated for primarily through these variables, and there are mass checks within the model that will reiterate a time step adjusting physically unrealistic values. As a result, water imbalance in the VICET simulations, when present, remains minimal. The remote sensing–based observation, default model simulation (VIC), and the simulation including ET bias correction (VICET) are compared in Fig. 3. The VICET simulation was run over the period 1980–2011, which is the span of the NLDAS-2 forcing data.

Fig. 3.

Comparison of LandFlux remotely sensed ET (Obs), default model ET (VIC), and bias-corrected ET used to force the ET-adjusted model (VICET).

Fig. 3.

Comparison of LandFlux remotely sensed ET (Obs), default model ET (VIC), and bias-corrected ET used to force the ET-adjusted model (VICET).

Last, a combination model (VICET–VEG) was also run for the time period with LAI data available, in which LAI has a minimal effect on ET because of the use of prescribed ET. LAI’s effects are more limited but still alter potential canopy storage, interception, infiltration, and root uptake. The VICVEG and VICET–VEG model simulations are run for the period 2003–11 because of the more limited availability of LAI data. Table 1 summarizes the four different experiments and the acronyms used to describe them.

Table 1.

Summary of experimental simulations and the acronyms used to describe each.

Summary of experimental simulations and the acronyms used to describe each.
Summary of experimental simulations and the acronyms used to describe each.

It is expected that after incorporating these data, the data-enhanced model will perform significantly better in simulating river discharge and its magnitude, seasonality, timing, soil moisture, temporal variation, and long-term trends. This will facilitate a more accurate understanding and attribution of past hydrological variability and changes. In addition, the difference between the default model results and the data-enhanced model results will help characterize the range of model-related uncertainties. To quantify the model improvement in simulating river flow attributed to the use of remote sensing data, a significance test was done to determine whether the correlation coefficient between model and observed river flow in an experiment is significantly different or higher than that in the default VIC run. The null hypothesis here is that the correlations are not different from one another, and Fisher’s test is used to determine the significance of the differences.

Assuming that the relationship between the remote sensing ET data and the default model simulation stays stationary, future simulations for both the default model (VIC) and the data-enhanced model VICET can be run for a future period 2046–65 using meteorological forcing derived from the NARCCAP models output. This allows for an examination of the impact of the ET bias correction on the simulated long-term trends of hydrological processes, including hydrological extremes such as peak and minimum discharge magnitude and timing, floods, drought, and water cycle fluxes.

3. Results

a. Model improvements

Because of data availability, the primary method for determining the model-simulated improvements is by comparing the various simulated discharges to observation. The river flow was simulated using the default model (VIC) as well as each data-enhanced version (VICET, VICVEG, and VICET–VEG). Because of the nature of the alterations made to the various model versions, the impact on model results is expected to manifest at different temporal scales. The prescribed ET is devised off of month-specific relationships between observation and simulation, and the relationships do not vary from year to year, whereas the MODIS LAI incorporated into the model is specifically designed to introduce interannual variations into the model. Our analyses are therefore conducted at several temporal scales, ranging from daily to interannual.

Compared to the default VIC simulation, the seasonal cycle of discharge in VICET (Fig. 4) shows great improvement. This specific analysis excludes the calibration period in which observational data were used to calibrate the model–observation ET relationship. The results therefore indicate how the relationships hold and whether they are capable of improving the model performance for time periods when observational data are not available. Figure 4 clearly displays that the dominant improvements made to the seasonal discharge occur during the summer and, to a lesser extent, the fall months. This is because of the greatest changes in ET between the default and VICET model occurring during these months. As the default model did best at estimating ET in the winter and spring months, there is not much room to further improve runoff simulations in those season. The correlations for Fig. 4, as well as correlation improvements on the daily, biweekly, and monthly time scales, can be found in Table 2. Because of the nature of the enhancements made to VICET, improvements were intended to manifest particularly on the seasonal scope. However, there are also substantial significant improvements on the monthly, biweekly, and even-finer daily time scale.

Fig. 4.

The (left) mean accumulated monthly discharge and (right) monthly bias from USGS observation at the (top) Thompsonville and (bottom) West Lebanon stations for the period of simulation excluding the availability of the ET data. The accumulated discharge is measured as the sum of all daily average cfs rates rather than converting to cubic feet.

Fig. 4.

The (left) mean accumulated monthly discharge and (right) monthly bias from USGS observation at the (top) Thompsonville and (bottom) West Lebanon stations for the period of simulation excluding the availability of the ET data. The accumulated discharge is measured as the sum of all daily average cfs rates rather than converting to cubic feet.

Table 2.

Comparison of streamflow for 1980–85 and 1996–2011 (for the period without observational ET data) in terms of the correlation and RMSE at Thompsonville. Boldface indicates significantly improved correlations at p = 0.05 in comparison to default model.

Comparison of streamflow for 1980–85 and 1996–2011 (for the period without observational ET data) in terms of the correlation and RMSE at Thompsonville. Boldface indicates significantly improved correlations at p = 0.05 in comparison to default model.
Comparison of streamflow for 1980–85 and 1996–2011 (for the period without observational ET data) in terms of the correlation and RMSE at Thompsonville. Boldface indicates significantly improved correlations at p = 0.05 in comparison to default model.

When comparing the improvements of all four model versions in a single graph (Fig. 5), it is clear that the VICVEG model does not show any marked improvement on the seasonal cycle time scale. In the summer months, when the impact of the ET bias correction is greatest, the dynamic LAI is lower on average than the climatological LAI used in the default model (Fig. 1). It is because of this reason that ET is decreased during these months, providing a greater overestimation of streamflow than with the default model. Perhaps more important than magnitude, the dynamic LAI does not appear to better capture the average seasonal variations. The results for the VICET–VEG combination model also very closely resemble those of the VICET model on the seasonal scale. In addition to improvement at the seasonal time scale, improvements on the biweekly time scale were also achieved (Fig. 6) for both the VICET and VICET–VEG models. Table 3 contains the correlations and root-mean-square errors (RMSEs) for every model version for the period of 2003–11. Whereas the daily, biweekly, and monthly time series are simply the accumulated values in each period, the seasonal cycle is the long-term average of each month’s accumulated discharge, and the interannual variability is computed based on the normalized anomalies of discharge in each month to remove the impact of seasonal cycle (i.e., subtracting the long-term mean of each month from the discharge time series and normalizing the anomalies with the standard deviation of that corresponding month). Table 3 exhibits just how important the time scale analyzed is toward the impact of data enhancement on improving streamflow estimates. Although the VICVEG model does not show any improvements on the finer time scales, it is more capable of enhancing the accuracy of streamflow on the interannual time scale (Fig. 6, Table 3) because of the introduction of the interannually varying LAI.

Fig. 5.

The mean accumulated monthly discharge at both stations for each model version. The period of analysis is the period of simulation shared by all datasets (2003–11).

Fig. 5.

The mean accumulated monthly discharge at both stations for each model version. The period of analysis is the period of simulation shared by all datasets (2003–11).

Fig. 6.

The (top) accumulated biweekly discharge for 2007–11 and (bottom) standardized anomalies of interannual monthly discharge (2003–11) for each model version at Thompsonville.

Fig. 6.

The (top) accumulated biweekly discharge for 2007–11 and (bottom) standardized anomalies of interannual monthly discharge (2003–11) for each model version at Thompsonville.

Table 3.

Comparison of streamflow for the period of simulation shared by all model versions (2003–11). Includes correlation and RMSE for both discharge stations. Boldface indicates significantly improved correlations at p = 0.05 in comparison to default model.

Comparison of streamflow for the period of simulation shared by all model versions (2003–11). Includes correlation and RMSE for both discharge stations. Boldface indicates significantly improved correlations at p = 0.05 in comparison to default model.
Comparison of streamflow for the period of simulation shared by all model versions (2003–11). Includes correlation and RMSE for both discharge stations. Boldface indicates significantly improved correlations at p = 0.05 in comparison to default model.

The combination model (VICET–VEG) shows improved streamflow estimates on every time scale analyzed, achieving the strongest correlations on the daily, biweekly, monthly, and seasonal cycle time scales. In addition to the interannual monthly analysis, the interannual variability of each season was examined separately (Table 4). The VICVEG model better captures variability, particularly in the winter and spring and not as much in the fall or in the summer months, despite many of the dynamic variations in LAI occurring in the winter and summer months.

Table 4.

Comparison of the seasonal [December–February (DJF), March–May (MAM), JJA, and September–November (SON)] interannual variability of streamflow (2003–11) at Thompsonville.

Comparison of the seasonal [December–February (DJF), March–May (MAM), JJA, and September–November (SON)] interannual variability of streamflow (2003–11) at Thompsonville.
Comparison of the seasonal [December–February (DJF), March–May (MAM), JJA, and September–November (SON)] interannual variability of streamflow (2003–11) at Thompsonville.

The improvement of model performance in simulating runoff and river flow indicates that the model performance in simulating soil moisture must have been improved too. The combination model VICET–VEG does indeed produce a seasonal cycle of soil moisture that better matches the variations of the observational seasonal cycle, including the summer season decrease (which is not as well captured by the default model). This is consistent with the improved seasonal cycle of discharge. However, there is disparity in terms of magnitude between simulation and remotely sensed soil moisture, as the model top soil layer is 10 cm thick while remote sensing soil moisture pertains to the top 2–3 cm of soil surface. Further comparisons are therefore not presented here as they are subject to larger uncertainties because of the depth-scale mismatch.

b. Hydrological changes and trends

After establishing that the data-enhanced models are capable of more accurately reconstructing past variability and magnitude of streamflow, it is important to examine whether the historic and future hydrological changes and long-term trends produced by the data-enhanced model may differ from those produced by the default model. This is especially the case for the VICET model since the impact of the ET data enhancement is not limited to the period when ET data are available and is therefore applicable to future predictions. Interestingly, it was found that the VICET-modeled future predictions differ greatly from the default model in not only the magnitude but also in some cases the direction of changes of water cycle variables, including some extreme indicators.

Daily peak and minimum discharge were analyzed in terms of both magnitude and timing for the Thompsonville station. Although there was minimal difference in the peak flow changes between the two models, the minimum discharge analysis shows drastic changes. Between our historic (1980–2011) and future (2046–65) simulations, the default model displays a change in the multiyear mean of the annual minimum daily flow from 2734 to 3443 ft3 s−1 (hereafter cfs; where 1 ft3 s−1 = 0.028 317 m3 s−1), equivalent to an average increase of 12 cfs yr−1. However, the VICET model produces a changing mean from 1592 to just 607 cfs, equivalent to an average decrease of 13 cfs yr−1. To provide some context, the default model daily mean flow decreased from 21 336 to 18 094 cfs between historic and future periods, a 15.2% relative decrease, and VICET daily mean flow decreased from 16 843 to 14 016 cfs, a 16.8% relative decrease. In addition to the opposite trends of the minimum flow magnitude, the mean timing of the daily minimums (which typically occurs in summer or fall) was delayed 22 days from 16 September to 8 October for VICET and was only delayed 14 days from 23 August to 6 September for the default model. Figure 7 shows the 5-day minimum discharge as well as the center of volume defined as the date at which half the discharge passes through the discharge station as an indication of seasonal timing. The 5-day minimum discharge has a similar trend to the absolute minimum discharge in each of the model simulations, with the two simulations producing opposite trends. The center of volume, which does not change much between past and future in the default model, shows an earlier date in VICET.

Fig. 7.

The (left) 5-day min accumulated discharge and (right) center of volume date at Thompsonville. The past (1980–2011) and future (2046–65) for both the default and ET-adjusted model versions are included.

Fig. 7.

The (left) 5-day min accumulated discharge and (right) center of volume date at Thompsonville. The past (1980–2011) and future (2046–65) for both the default and ET-adjusted model versions are included.

The future flood risk analysis conducted in this study is based on river discharge using the same methodology as Parr and Wang (2014). The analysis compares past and future simulations by adjusting VIC-simulated discharge based on the correspondence found between observational and VIC-simulated discharge derived from the past climate and a rating curve derived from observational data. Specifically, a rating curve was first derived based on past observational discharge and gauge height for the Thompsonville station from data provided by USGS. Second, the correspondence between observational and VIC-simulated discharge was derived from the past climate. The derived gauge height data were then used to analyze the flooding risks with reference to the “action level” gauge height (the height at which flooding becomes a risk and actions should be taken) for Thompsonville determined by the NOAA/National Weather Service Advanced Hydrologic Prediction Service (NWS 2013). The flood analysis showed a fairly dramatic future change in characteristics when analyzed based on output from the default VIC (Parr et al. 2015, manuscript submitted to Global Planet. Change). The changes include an increasing duration of floods with a decreasing frequency, leading to a similar number of total flood days. Output from the ET-enhanced model produces the same change in characteristics, with no qualitative difference in the future change signal between VIC and VICET (Fig. 8).

Fig. 8.

The (top) number of flood days, (middle) mean length of flood events, and (bottom) number of events for the past (1980–2011) and future (2046–65). Displayed for the future is the mean of all three future datasets for the default and ET-adjusted model versions.

Fig. 8.

The (top) number of flood days, (middle) mean length of flood events, and (bottom) number of events for the past (1980–2011) and future (2046–65). Displayed for the future is the mean of all three future datasets for the default and ET-adjusted model versions.

A drought analysis was also conducted that measures the change in drought signals between the future and historic scenarios for both model versions. It is adapted from a similar one conducted by Sheffield and Wood (2007), where their percentile was based on monthly rather than daily data. In this study, for each grid cell in the basin, a month-specific 20th-percentile daily soil moisture threshold was identified based on historic (1980–2011) data. If soil moisture was below this threshold for 15 or more days in a specific month, it was considered a drought month for that specific grid cell. The mean duration of droughts and the number of short- (4–6 months), medium- (7–11 months), and long-term (≥12 months) droughts were then identified. Figure 9 displays a clear and fairly drastic difference in trends between the two model versions. The default VIC produces little to no change in drought for any of the categories, but VICET predicts universal increases. The influence of model version is evident when comparing the difference in trends shown in the right-most column. The future increase of the mean drought duration changes from 0.21 months for the default VIC to 4.48 months for VICET. The future increase in the quantity of the droughts changes from 0.72 for the default VIC to 12.66 for VICET for the number of short-term droughts per 20 years, from 0.80 to 5.12 for the number of medium-term droughts per 20 years, and from 0.54 to 1.80 for the number of long-term droughts per 20 years. The VICET future scenario is capable of having such a large number of drought hits because of the percentile threshold used to define droughts being taken from the default historic scenario. Because of an increase of ET in VICET compared with VIC and an increase of ET in the future scenarios, the soil moisture is substantially less in the VICET future than in the VIC historic scenario, which the drought threshold was based upon.

Fig. 9.

Analysis of drought trends for the past (1980–2011) and future (2046–65) based on soil moisture: (from top to bottom) the mean duration of droughts (months), the number of short-term droughts, the number of medium-term droughts, and the number of long-term droughts for (left) change (future minus past) for the default model, (middle) change (future minus past) for the ET-adjusted model, and (right) the difference between model versions.

Fig. 9.

Analysis of drought trends for the past (1980–2011) and future (2046–65) based on soil moisture: (from top to bottom) the mean duration of droughts (months), the number of short-term droughts, the number of medium-term droughts, and the number of long-term droughts for (left) change (future minus past) for the default model, (middle) change (future minus past) for the ET-adjusted model, and (right) the difference between model versions.

To understand the results shown in Figs. 7 and 9, the differences between the two model versions in water cycle variables and their change between the historic and future scenarios (Fig. 10) are explored. There exists no observable difference between the two model versions in seasonal trends from the past to future for direct surface runoff or subsurface runoff, including groundwater and interflow. However, as one might expect from the response of drought characteristics, a larger magnitude of increase in future ET is produced by the VICET model, specifically in the summer season when droughts are most likely to occur, and a decrease of summer soil moisture in VICET is not simulated in the default VIC.

Fig. 10.

Seasonal change in (a) ET and (b) soil moisture between past (1980–2011) and future (2046–65) for both default and ET-adjusted model versions.

Fig. 10.

Seasonal change in (a) ET and (b) soil moisture between past (1980–2011) and future (2046–65) for both default and ET-adjusted model versions.

4. Summary and discussion

Incorporating remote sensing data into VIC has made a measureable improvement toward recreating historic river flow estimates in the Connecticut River basin. The improvements indicate that incorporation of remote sensing information into hydrologic models can improve model accuracy and insight into current and future hydrologic processes and can help characterize the model-related uncertainties in hydrological predictions.

Hydrologic modeling can be affected by four main sources of uncertainty (Renard et al. 2010), including input uncertainty arising from sampling or measurement errors, output uncertainty related to analysis such as with rating curves, model uncertainties arising from the oversimplification or representation of hydrological processes, and existing parametric uncertainty reflecting the inability to specify exact values of model parameters for a variety of reasons (Renard et al. 2010). In this study, we have attempted to minimize the input uncertainty involved in LAI with more accurate and specific data and replaced model uncertainty in representing the ET process with input uncertainties related to ET data. Remote sensing data are chosen for this study because of their high spatial coverage, as the ultimate goal is to improve model estimates within basins without accurate streamflow or ground-based measurements. Potential uncertainties related to the choice of the ET data can be examined based on the relationship between the simulated and observed datasets at the gridcell level. This relationship is very similar between the LandFlux dataset used in the study and the MODIS ET dataset of Mu et al. (2007). The LandFlux data were chosen for use here over the MODIS ET data for better spatial coverage, as the MODIS ET data are missing for some grid cells within our model domain. The relationship between the default VIC ET and remote sensing ET data applied in this study is based on a simple ratio bias adjustment. The bias-correction algorithm based on this month- and gridcell-specific ratio leads to adjustment to both the mean and variance of the model ET, but a majority of the differences in model performance and in future projections are due to adjustment of the mean ET. As such, the use of an alternative, more sophisticated bias-correction approach is not expected to result in qualitative changes in the results of this study. As any gridded meteorological forcing data contain a certain degree of bias, the ET bias correction implemented here corrects for not only model bias but also potential forcing errors. This may contribute to uncertainties in the application of the methodology to future predictions.

As for the choice of LAI datasets, the MODIS data provide the most accurate estimates available with good spatial coverage and fine spatial resolution, and they were used in the majority of similar studies examining the impact of including dynamic LAI (Wattenbach et al. 2012; Tang et al. 2012; Zhou et al. 2013; Ford and Quiring 2013). For the LAI and fraction of photosynthetically active radiation (fPAR) data retrievals, there have been algorithm refinements targeted to be consistent with field measurements over every type of biome, but with particular focus on woody vegetation, which is the most prominent biome found in the Connecticut River basin.

In conclusion, the incorporation of the remotely sensed data is capable of creating statistically measureable model improvements, particularly toward the estimations of river flow. Because of the nature of the alterations made to the various model versions, improvements manifest on different temporal scales. As the prescribed ET is devised off of monthly relationships of daily ET at the gridcell level between observation and simulation, it is the seasonal signal that shows the greatest improvements in streamflow. The MODIS LAI incorporated into the model is specifically designed to introduce interannual variations into the model, so it is not surprising that it is on this particular scale that we see its performance enhanced the most. However, more notable are the VICET model version improvements that occur on a variety of time scales, including improved estimates of streamflow variations on the finer daily and biweekly scales. The observational soil moisture dataset from the ESA that uses active and passive microwave data is only able to capture the top 10 cm of the soil. The difficulty to retrieve soil moisture from remote sensing likely also contributes to the differences between simulations and observation.

Another interesting finding from this study was how the VICET–VEG combination model was able to find a balance between the other two enhanced versions displaying the most significant increases on the seasonal, monthly, biweekly, and daily scales, as well as still slightly improving the performance at the interannual time scale. As it incorporates the prescribed ET, it was possible that the effects of the dynamic LAI would be marginalized, as one of LAI’s greatest impacts is on transpiration and the surface energy budget.

The VICVEG model better captures variability, particularly in the winter and spring and not as much in the fall or in the summer months. The spring season improvements could be accounted for by changing greening onset dates that are particularly important to the hydrology of this region during these months. The winter seasons display some of the greatest variations from year to year in terms of LAI, but so do the summer months, so this lack of improvement is difficult to account for, but it may be due to ET being more restricted by water availability and thus less dependent on the effects LAI would have on transpiration.

The ET ratio is applied to all five components of ET equally, and it is possible this may result in changes where and when such changes are not needed. For example if PET is estimated reasonably well, the canopy evaporation may be accurate in VIC, and only the transpiration may need adjustment because of limiting soil moisture. Similarly, the summer ET needs correcting in VIC, but winter and spring ET may not. These types of issues may be playing a role in why the revised ET reduces discharge magnitude accuracy in spring but improves in the summer. It is possible that the seasonal ET errors in VIC are associated with limiting soil water rather than PET estimation.

In the application of the ET bias correction to future predictions, compared with the default model, VICET produces a very different signal of future changes in soil moisture drought and minimum discharge; little difference is found in flooding characteristics or maximum discharge. This is because of the extent to which ET is altered in the months these extremes occur, and the extent of the bias correction’s seasonal influence may be region or model dependent. The analyses from both the default and VICET model versions point toward the future of the Connecticut River basin being characterized by fewer but also longer flood events. Note that VIC is not designed to predict floods, so these conclusions are based on an observational rating curve, a stage height threshold, and an adjustment of simulated discharge according to a linear relationship with observation (Parr et al. 2015, manuscript submitted to Global Planet. Change), and as such may be tenuous. The main point here is that there is no observable difference between model versions in the predicted future flooding risks. However, when compared to VIC, VICET does simulate more frequent and longer droughts. This drought signal was detected in soil moisture, as well as the absolute minimum and 5-day minimum streamflows.

Although VICET simulated more accurate historic streamflow, it cannot be said conclusively that the predictions of the enhanced model are more accurate in all respects. However, since the enhanced model produces streamflow with improved accuracy during time periods without ET observational data available, it is reasonable to expect that the ET enhancement may improve future predictions in the same manner.

Although this study focuses solely on the Connecticut River basin, the methodology developed to incorporate remote sensing data into physically based hydrological models is applicable to other regions for the same model and to other hydrological or land surface models as well. In fact, high spatial coverage is one of the greatest advantages of remote sensing data along with its high spatial resolution, which makes remotely sensed data particularly useful in areas of spare ground-based measurements. In the case of streamflow, this occurs when river gauges are not maintained, in regions where water is anthropogenically diverted, or where the natural flow has been altered. Depending on the deficiencies any particular model may have in a particular region, the incorporation of remote sensing data could provide a variety of improvements. Similar techniques could be useful in more complex models, or if applied by systems like NLDAS and GLDAS, could help to better improve our understanding of climatic or hydrologic change and perhaps even improve upon our future predictions.

Acknowledgments

The authors would like to acknowledge the funding support made available through a fellowship to Dana Parr from the NASA/Connecticut Space Grant Consortium (Grant P-743). We thank Elizabeth Clark and two anonymous reviewers for their constructive comments on an earlier version of the paper.

REFERENCES

REFERENCES
Ahmed
,
K. F.
,
G.
Wang
,
J.
Silander
,
A.
Wilson
,
J.
Allen
,
R.
Horton
, and
R.
Anyah
,
2013
:
Statistical downscaling and bias correction of climate model outputs for climate change impact assessment in the U.S. Northeast
.
Global Planet. Change
,
100
,
320
332
, doi:.
Bjerklie
,
D.
,
T.
Trombley
, and
R.
Viger
,
2011
:
Simulations of historical and future trends in snowfall and groundwater recharge for basins draining to Long Island Sound
.
Earth Interact.
,
15
, doi:.
Bormann
,
H.
,
2011
: Sensitivity analysis of 18 different potential evapotranspiration models to observed climatic change at German climate stations. Climatic Change, 104, 729–753, doi:.
Caya
,
D.
, and
R.
Laprise
,
1999
:
A semi-implicit semi-Lagrangian regional climate model: The Canadian RCM
.
Mon. Wea. Rev.
,
127
,
341
362
, doi:.
Dorigo
,
W. A.
,
K.
Scipal
,
R. M.
Parinussa
,
Y. Y.
Liu
,
W.
Wagner
,
R. A. M.
de Jeu
, and
V.
Naeimi
,
2010
:
Error characterization of global active and passive microwave soil moisture data sets
.
Hydrol. Earth Syst. Sci.
,
14
,
2605
2616
, doi:.
Eitzinger
,
J.
,
D.
Marinkovic
, and
J.
Hösch
,
2002
: Sensitivity of different evapotranspiration calculation methods in different crop–weather models. Integrated Assessment and Decision Support: Proceedings of the First Biennial Meeting of the iEMSs, Vol. 2, A. E. Rizzoli and A. J. Jakeman, Eds., 395–400. [Available online at www.iemss.org/iemss2002/proceedings/pdf/volume%20due/161_eitzinger.pdf.]
Fisher
,
J.
,
K.
Tu
, and
D.
Baldocchi
,
2008
:
Global estimates of the land–atmosphere water flux based on monthly AVHRR and ISLSCP-II data, validated at 16 FLUXNET sites
.
Remote Sens. Environ.
,
112
,
909
919
, doi:.
Fisher
,
J.
, and Coauthors
,
2009
:
The land–atmosphere water flux in the tropics
.
Global Change Biol.
,
15
,
2694
2714
, doi:.
Ford
,
T.
, and
S.
Quiring
,
2013
:
Influence of MODIS-derived dynamic vegetation on VIC-simulated soil moisture in Oklahoma
.
J. Hydrometeor.
,
14
,
1910
1921
, doi:.
Immerzeel
,
W. W.
, and
P.
Bold
,
2008
:
Calibration of a distributed hydrological model based on satellite evapotranspiration
.
J. Hydrol.
,
349
,
411
424
, doi:.
Kite
,
G. W.
, and
P.
Droogers
,
2000
:
Comparing evapotranspiration estimates from satellites, hydrological models, and field data
.
J. Hydrol.
,
229
,
3
18
, doi:.
Liang
,
X.
,
D. P.
Lettenmaier
,
E. F.
Wood
, and
S. J.
Burges
,
1994
:
A simple hydrologically based model of land surface water and energy fluxes for GSMs
.
J. Geophys. Res.
,
99
,
14 415
14 428
, doi:.
Liu
,
Y. Y.
,
W. A.
Dorigo
,
R. M.
Parinussa
,
R. A. M.
de Jeu
,
W.
Wagner
,
M. F.
McCabe
,
J. P.
Evans
, and
A. I. J. M.
van Dijk
,
2012
:
Trend-preserving blending of passive and active microwave soil moisture retrievals
.
Remote Sens. Environ.
,
123
,
280
297
, doi:.
Lu
,
J.
,
G.
Sun
,
S. G.
McNulty
, and
D. M.
Amatya
,
2005
:
A comparison of six potential evapotranspiration methods for regional use in the southeastern United States
.
J. Amer. Water Resour. Assoc.
,
41
,
621
633
, doi:.
Marshall
,
E.
, and
T.
Randhir
,
2008
:
Effect of climate change on watershed system: A regional analysis
.
Climatic Change
,
89
,
263
280
, doi:.
McKenney
,
M. S.
, and
N. J.
Rosenberg
,
1993
:
Sensitivity of some potential evapotranspiration estimation methods to climate change
.
Agric. For. Meteor.
,
64
,
81
110
, doi:.
Mearns
,
L.
, and Coauthors
,
2012
:
The North American Regional Climate Change Assessment Program: Overview of phase I results
.
Bull. Amer. Meteor. Soc.
,
93
,
1337
1362
, doi:.
Mu
,
Q.
,
F.
Heinsch
,
M.
Zhao
, and
S.
Running
,
2007
:
Development of a global evapotranspiration algorithm based on MODIS and global meteorology data
.
Remote Sens. Environ.
,
111
,
519
536
, doi:.
Naeimi
,
V.
,
K.
Scipal
,
Z.
Bartalis
,
S.
Hasenauer
, and
W.
Wagner
,
2009
:
An improved soil moisture retrieval algorithm for ERS and MetOp scatterometer observations
.
IEEE Trans. Geosci. Remote Sens.
,
47
,
1999
2013
, doi:.
NWS
,
2013
: Connecticut River at Thompsonville hydrograph. Advanced Hydrologic Prediction Service, NOAA/NWS. Accessed 12 September 2013. [Available online at http://water.weather.gov/ahps2/hydrograph.php?wfo=box&gage=tmvc3&hydro_type=2.]
Ottlé
,
C.
, and
D.
Vidal-Madjar
,
1994
:
Assimilation of soil moisture inferred from infrared remote sensing in a hydrological model over the HAPEX-MOBILHY region
.
J. Hydrol.
,
158
,
241
264
, doi:.
Pal
,
J. S.
, and Coauthors
,
2007
:
Regional climate modeling for the developing world: The ICTP RegCM3 and RegCNET
.
Bull. Amer. Meteor. Soc.
,
88
,
1395
1409
, doi:.
Parr
,
D. T.
, and
G. L.
Wang
,
2014
:
Hydrological changes in the U.S. Northeast using the Connecticut River basin as a case study: Part 1. Modeling and analysis of the past
.
Global Planet. Change
,
122
,
208
222
, doi:.
Renard
,
B.
,
D.
Kavetski
,
G.
Kuczera
,
M.
Thyer
, and
S.
Franks
,
2010
:
Understanding predictive uncertainty in hydrologic modeling. The challenge of identifying input and structural errors
.
Water Resour. Res.
,
46
,
W05521
, doi:.
Reynolds
,
C. A.
,
T. J.
Jackson
, and
W. J.
Rawls
,
2000
: Estimating soil water-holding capacities by linking the Food and Agriculture Organization soil map of the world with global pedon databases and continuous pedotransfer functions. Water Resour. Res., 36, 3653–3662, doi:.
Sheffield
,
J.
, and
E.
Wood
,
2007
:
Characteristics of global and regional drought, 1950–2000: Analysis of soil moisture data from off-line simulation of the terrestrial hydrological cycle
.
J. Geophys. Res.
,
112
,
D17115
, doi:.
Sheffield
,
J.
,
B.
Livneh
, and
E. F.
Wood
,
2007
: Representation of terrestrial hydrology and large-scale drought of the continental United States from the North American Regional Reanalysis. J. Hydrometeor., 13, 856–876, doi:.
Tang
,
Q.
,
E. R.
Vivoni
,
F.
Muñoz-Arriola
, and
D. P.
Lettenmaier
,
2012
:
Predictability of evapotranspiration patterns using remotely sensed vegetation dynamics during the North American monsoon
.
J. Hydrometeor.
,
13
,
103
121
, doi:.
Trambauer
,
P.
,
E.
Dutra
,
S.
Maskey
,
M.
Werner
,
F.
Pappernberger
,
L. P. H.
van Beek
, and
S.
Uhlenbrook
,
2014
:
Comparison of different evaporation estimates over the African continent
.
Hydrol. Earth Syst. Sci.
,
18
,
193
212
, doi:.
USGS
,
2011
: HYDRO1k documentation. Accessed 22 June 2015. [Available online at http://webgis.wr.usgs.gov/globalgis/metadata_qr/metadata/hydro1k.htm.]
USGS
,
2012
: USGS water data for Connecticut. Accessed 30 January 2012. [Available online at http://waterdata.usgs.gov/ct/nwis.]
Vano
,
J.
,
T.
Das
, and
D.
Lettenmaier
,
2012
:
Hydrologic sensitivities of Colorado River runoff to changes in precipitation and temperature
.
J. Hydrometeor.
,
13
,
932
949
, doi:.
Vinukollu
,
R.
,
E.
Wood
,
C.
Ferguson
, and
J.
Fisher
,
2011
:
Global estimates of evapotranspiration for climate studies using multi-sensor remote sensing data: Evaluation of three process-based approaches
.
Remote Sens. Environ.
,
115
,
801
823
, doi:.
Wattenbach
,
M.
,
D.
Franz
,
W.
Liang
,
M.
Schmidt
,
F.
Seitz
, and
A.
Güntner
,
2012
: Integration of MODIS LAI products into the hydrological model WGHM indicate the sensitivity of total water storage simulations to vegetation cover dynamics. Geophysical Research Abstracts, Vol. 14, Abstract EGU2012-10116. [Available online at http://meetingorganizer.copernicus.org/EGU2012/EGU2012-10116.pdf.]
Weiß
,
M.
, and
L.
Menzel
,
2008
:
A global comparison of four potential evapotranspiration equations and their relevance to stream flow modelling in semi-arid environments
.
Adv. Geosci.
,
18
,
15
23
, doi:.
Wood
,
A. W.
,
E. P.
Maurer
,
A.
Kumar
, and
D. P.
Lettenmaier
,
2002
:
Long-range experimental hydrologic forecasting for the eastern United States
.
J. Geophys. Res.
,
107
,
4429
, doi:.
Wood
,
A. W.
,
L. R.
Leung
,
V.
Sridhar
, and
D. P.
Lettenmaier
,
2004
:
Hydrological implications of dynamical and statistical approaches to downscaling climate model surface temperature and precipitation fields
.
Climatic Change
,
62
,
189
216
, doi:.
Xia
,
Y.
,
M.
Ek
, and
H.
Wei
,
2012a
:
Comparative analysis of relationships between NLDAS-2 forcings and model outputs
.
Hydrol. Processes
,
26
,
467
474
, doi:.
Xia
,
Y.
, and Coauthors
,
2012b
: Continental-scale water and energy flux analysis and validation for the North American Land Data Assimilation System project phase 2 (NLDAS-2): 1. Intercomparison and application of model products. J. Geophys. Res., 117, D03109, doi:.
Xia
,
Y.
, and Coauthors
,
2012c
: Continental-scale water and energy flux analysis and validation for North American Land Data Assimilation System project phase 2 (NLDAS-2): 2. Validation of model-simulated streamflow. J. Geophys. Res., 117, D03110, doi:.
Xia
,
Y.
,
M.
Hobbins
,
Q.
Mu
, and
M.
Ek
,
2015
:
Evaluation of NLDAS-2 evapotranspiration against tower flux site observations
.
Hydrol. Processes
, 29, 1757–1771, doi:.
Zhang
,
Y. Q.
,
N. R.
Viney
,
F. H. S.
Chiew
,
A. I. J. M.
van Dijk
, and
Y. Y.
Liu
,
2011
: Improving hydrological and vegetation modelling using regional calibration schemes together with remote sensing data. MODSIM2011: 19th International Congress on Modelling and Simulation, F. Chan, D. Marinova, and R. S. Anderssen, Eds., Modelling and Simulation Society of Australia and New Zealand, 3448–3454. [Available online at http://mssanz.org.au/modsim2011/I4/zhang.pdf.]
Zhou
,
Y.
,
Y.
Zhang
,
J.
Vaze
,
P.
Lane
, and
S.
Xu
,
2013
:
Improving runoff estimates using remote sensing vegetation data for bushfire impacted catchments
.
Agric. For. Meteor.
,
182-183
,
332
341
, doi:.