Continental satellite soil moisture data assimilation improves root-zone moisture analysis for water resources assessment

Journal of Hydrology 519 (2014) 2747–2762

Contents lists available at ScienceDirect

Journal of Hydrology journal homepage: www.elsevier.com/locate/jhydrol

Continental satellite soil moisture data assimilation improves root-zone moisture analysis for water resources assessment L.J. Renzullo a,⇑, A.I.J.M. van Dijk b, J.-M. Perraud a, D. Collins c, B. Henderson d, H. Jin d, A.B. Smith e, D.L. McJannet f a

CSIRO Land and Water, Canberra, ACT, Australia Australian National University, Fenner School of Environment and Society, Canberra, ACT, Australia CSIRO Advanced Scientific Computing, Kensington, WA, Australia d CSIRO Computational Informatics, Canberra, ACT, Australia e Bureau of Meteorology, Melbourne, Vic., Australia f CSIRO Land and Water, Brisbane, Qld, Australia b c

a r t i c l e

i n f o

Article history: Available online 11 August 2014 Keywords: Satellite soil moisture Data assimilation Water resources Cosmic ray sensor

s u m m a r y A framework was developed for the continental assimilation of satellite soil moisture (SM) into an operational water balance modelling system. The ensemble Kalman filter (EnKF) was implemented to assimilate AMSR-E and ASCAT-derived SM products into the landscape model of the Australian Water Resources Assessment system (AWRA-L) and generate ensembles of daily top-layer and shallow rootzone soil moisture analyses for the continent at 0.05° resolution. We evaluated the AWRA-L SM estimates with and without assimilation against in situ moisture measurements in southeast Australia (OzNet), as well as against a new network of cosmic-ray moisture probes (CosmOz) spread across the country. Results show that AWRA-L root-zone moisture estimates are improved though the assimilation of satellite SM: model estimates of 0–30 cm moisture content improved for more than 90% of OzNet sites, with an increase in average correlation from 0.68 (before assimilation) to 0.73 (after assimilation); while estimates 0–90 cm moisture improved for 60% of sites with increased average correlation from 0.56 to 0.65. The assimilation of AMSR-E and ASCAT appeared to yield similar performance gains for the top-layer, however ASCAT data assimilation improved root-zone estimation for more sites. Poor performance of one data set was compensated by the other through joint assimilation. The most significant improvements in AWRA-L root-zone moisture estimation (with increases in correlation as high as 90%) occurred for sites where both the assimilation of satellite soil moisture improved top-layer SM accuracy and the open-loop deep-layer storage estimates were reasonably good. CosmOz SM measurements exhibited highest correlation with AWRA-L estimates for modelled root-zones layer thicknesses ranging from 20 cm to 1 m. Slight improvements through satellite data assimilation were observed for only 2 of 7 CosmOz sites, but the comparison was affected by a short data overlap period. The location of some of the CosmOz probes was not optimal for evaluation of satellite SM assimilation, but their utility is demonstrated and the observations may become suitable for assimilation themselves in future. Ó 2014 Elsevier B.V. All rights reserved.

1. Introduction Over the past decade many researchers have demonstrated that the assimilation of remotely-sensed soil moisture (SM) products into land surface models can improve soil water balance predictions which can in turn lead to improved estimates of evaporative fluxes, drainage and runoff (e.g. Reichle and Koster, 2005; Brocca ⇑ Corresponding author. Address: CSIRO Land and Water, Black Mountain Laboratories, Canberra, ACT 2601, Australia. Tel.: +61 2 6246 5758. E-mail address: [email protected] (L.J. Renzullo). http://dx.doi.org/10.1016/j.jhydrol.2014.08.008 0022-1694/Ó 2014 Elsevier B.V. All rights reserved.

et al., 2010; Dharssi et al., 2011; Draper et al., 2011; Pipunic et al., 2013). The ability to constrain water balance estimation over large areas offers great potential for continental water resource assessment, particularly in parts of the landscape where traditional ground observation networks have sparse or intermittent coverage. Conventional use of satellite observations for water balance studies is via the direct or indirect measurement of key components of the water cycle, including precipitation (e.g. Joyce et al., 2004; Huffman et al., 2007; Hou et al., 2008), evapotranspiration (e.g. Anderson et al., 2007; Mu et al., 2007; Kalma et al., 2008; Guerschman et al., 2009), surface soil moisture (Wagner et al.,

2748

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

1999; Njoku et al., 2003; Owe et al., 2008), terrestrial water storage through gravitational anomalies (e.g. Rodell et al., 2006, 2009; Van Dijk et al., 2011), river water storage and runoff (Alsdorf and Lettenmaier, 2003; Berry et al., 2005). However, observation-driven attempts to attain water balance closure with these remotely-sensed data have identified errors and biases in the individual data sets, as well as inconsistencies between linked variables, as limiting factors to closure (e.g. McCabe et al., 2008; Sheffield et al., 2009; Gao et al., 2010; Pan et al., 2012). A water balance modelling system constrained with multiple remotely-sensed and ground-based observations would maximise consistency (closure) among water balance variables and in principle result in improved estimation (particularly in ungauged areas) provided the data are integrated into the modelling system in ways that incorporate both model and observation uncertainties. Data assimilation is one approach to model-data integration that allows optimal combination of model with observations, and improved estimation, provided respective errors are adequately specified. To-date, however, there are very few, if any, such data assimilation systems that are capable of exploiting satellite remote sensing data for operational water resources assessment. Van Dijk and Renzullo (2011) argue that this may be due to the small number of operational missions (and hence data continuity issues) and the generally coarse spatial scale of the satellite data and models used – this is the impetus of the grand challenge of hyperresolution (1 km) satellite remote sensing proposed for water resource monitoring by Wood et al. (2011). The Australian Water Resources Assessment (AWRA) system is a high-resolution modelling framework for continental water resources assessment and accounting with model-data integration at the core of its development. Unlike most global land surface models, the AWRA system runs at a fine (5-km) resolution, is extensively calibrated with surface networks of streamflow, and is driven by high-quality meteorological surfaces derived from surface observations. The AWRA system currently operational in the Bureau of Meteorology does have a component specifically for the assimilation of satellite data into the system, but this functionality remains dormant pending further development and testing. The work presented here describes an aspect of the testing of this data assimilation component. Specifically, we investigated whether the assimilation of satellite soil moisture products into the landscape water balance model component of the AWRA system (AWA-L) leads to improved moisture estimation in both the top (near surface) layer and in the shallow (<1 m) root zone. The AWRA-L model (described in Section 2.1) is part of a larger modelling system designed to augment surface metering where available and provide comprehensive coverage of key water balance terms for national scale water resources assessment and accounting. The method of satellite data assimilation employed (described in Section 2.2) is a popular and demonstrably effective technique for sequential updating of model soil moisture states where and when satellite data are available. Evaluating root-zone soil moisture estimation may be considered one of the first steps towards quality assurance of the model performance, as soil moisture is a key variable in the partitioning of rainfall into evaporation, infiltration and runoff, as well as being a quantity of interest (e.g. for drought assessment) in its own right that is reported in annual Australian Water Resources Assessments (Bureau of Meteorology, 2013). While both the satellite soil moisture data used (described in Section 3.1), the data preprocessing and error characterisation (triple collocation, Section 3.2), and the water balance modelling are continental in extent, our evaluations focuses on an in situ monitoring network in southeastern Australia (Section 3.3). In addition, this is the first study to evaluate assimilation results against the new cosmic-ray soil moisture probe that is being deployed across Australia. We present (in

Section 4) the results of the evaluation of the assimilation results against both ground-based soil moisture monitoring networks, and assess improvement against the unconstrained (open-loop) model estimates. 2. Method 2.1. The Australian Water Resources Assessment landscape model: AWRA-L The Australian Water Resources Assessment (AWRA) system was developed by Australia’s Commonwealth Scientific and Industrial Research Organisation (CSIRO) and Bureau of Meteorology (BoM). The system is comprised of 3 model components: the landscape (AWRA-L), river (AWRA-R) and groundwater (AWRA-G) models. The system generates the water balance terms that underpin the BoM’s mandatory reporting on the state of Australia’s water resources and is under continual development (Van Dijk and Renzullo, 2011; Stenson et al., 2011). For the investigation presented here, we limited our focus to the landscape model AWRA-L (Van Dijk, 2010). AWRA-L describes the temporal evolution of water stores and fluxes across a region of interest. Gridded forcing data (spatial estimates of meteorological observations, see Section 2.2) drive the model to produce spatial water balance estimates on a grid of 0.05°  0.05° cells across Australia. Note that in this system component each cell is modelled independently of its neighbours (i.e. there is no lateral transport of water). Fig. 1 provides a schematic for the AWRA-L model representation of the soil column. Fig. 1a shows the unsaturated zone partitioned into three conceptual soil water stores: the first corresponding to the uppermost soil layer, called the top layer; the second corresponding to the part of the soil where water is extracted by shallow-rooted vegetation, referred to as the shallow root layer; and the third is where deep-rooted vegetation extract water, referred to the deep root layer. Each layer is characterised by a maximum water holding capacity (field capacity) parameter, denoted S0FC, SsFC and SdFC for the top-layer, shallow and deep root soil layers respectively. Unlike many land surface models, AWRA-L specifies only the water storage, Sz, in the soil layers rather than prescribing a layer thickness and porosity (and hFC and hWP ) based on pedotransfer functions. This was done to avoid the model parameter estimation equifinality issues (Beven and Freer, 2001) associated with introducing mathematically equivalent parameters into the AWRA-L model. The free parameters with regard to soil layer specification are the field capacity values (i.e. S0FC, SsFC and SdFC); the water availability in Fig. 1b is only used to provide indicative soil layer thicknesses and to aid in the evaluation of the AWRA-L storage estimates against ground data. Viney et al. (2011) describe the method for estimating these field capacity values (among other model parameters) for application of the AWRA system across Australia. Here we have used the values specific to the currently operational version 3.0 of AWRA-L. The physical thicknesses of the soil layers, Z (mm), can be approximated as,

Z¼

SzFC ðhFC  hWP Þ

ð1Þ

where SzFC is field capacity water storage (in mm) for soil layer z, and the difference between field capacity and wilting point for each soil layer in volumetric units, hFC  hWP , represents the soil available water. Spatial estimates of soil water availability from the Australian Soil Resource Information System (Fig. 1b) (http://www.asris. csiro.au) are used to provide indicative estimates of soil layer thicknesses for the top, shallow-root and deep-root layers following

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

2749

Fig. 1. Soil column representation in the AWRA-L model: (a) schematic of the pathways of water movement between layers and in and out of the soil column; (b) available water estimates for Australia from digital soil maps (ASRIS); and (c) histogram of estimated AWRA-L soil layer thickness value across Australia, derived from model calibrated field capacity water storage parameters for the top (S0FC), shallow root (SsFC) and deep root (SdFC) soil layers and soil map available water.

the relationship in Eq. (1) (Fig. 1c). The top-soil layer in AWRA-L is the closest model equivalent to the emitting layer seen by C- or Lband remote sensing systems, and has thicknesses across Australia varying between 7 and 9 cm. The shallow- and deep-root layers have estimated thicknesses between 10–21 cm and 6–10 m, respectively. Therefore a shallow root-zone for the AWRA-L model can be broadly defined as the top 0–30 cm of the soil column, and a deep root-zone the top 0–10 m of soil. Water enters the soil column through the top layer as net precipitation (i.e. precipitation minus rainfall interception) and into the deep root layer through capillary rise from groundwater. Water leaves the column through soil evaporation (which is estimated as part of the landscape water balance), extraction by shallow- and deep-rooted vegetation, or drainage into the groundwater store (a portion of which contributes to modelled river runoff through discharge). Drainage between layers is calculated in AWRA-L as:

DZ ¼ K FC exp½bð1  SZ =SzFC ÞSZ ;

ð2Þ

where SZ is the amount of water stored in layer z and parameters KFC and b are estimated a priori through the calibration of AWRA-L against streamflow measurements (Viney et al., 2011). Van Dijk and Marvanek (2010) derived Eq. (2) after exploring several alternative soil water percolation models and found that it simulates soil moisture dynamics with very comparable accuracy to the more computationally expensive and sophisticated (Richards’ equation) while producing better agreement with daily streamflow observations in 198 Australian catchments than alternative formulations. 2.2. Data assimilation method In this study we chose the ensemble Kalman filter (EnKF, Evensen, 2003) approach to assimilate satellite soil moisture into AWRA-L. The EnKF is a relatively simple, flexible and therefore popular technique for assimilating satellite data into land surface models (e.g. Reichle et al., 2002, 2008; Crow and Wood, 2003; Crow and Van den Berg, 2010; Draper et al., 2012; Kumar et al., 2008, 2009; Pipunic et al., 2008, 2013). It is suitable for use with moderately nonlinear systems and uses ensembles of model states to derive error statistics needed to optimally combine models with noisy observations. Ensembles of model states are generated

during the so-called forecast step by propagating the model from the previous to the current time step with meteorological data and perturbations (more in Section 2.2.1). The ensembles of forecast states are used to compute model error variances. The analysis or update step of the EnKF involves adjusting the forecast estimates towards available observations by an amount that is weighted by model and observation error variances (Section 2.2.2). Kumar et al. (2009) identify the two main mechanisms through which satellite soil moisture can impart constraint on moisture stores in deeper soil layers: the first is a result of model physics, where improvements in one soil layer is transferred to the next (and so on) through percolation processes and water balance (during the forecast step); the second is through simultaneous adjustment of multiple soil layers (model states) given observations and the variance–covariance structures between model states and observations (during the analysis step). In our application of the EnKF, the states of the system, xt, were the AWRA-L model storages in the top, shallow root and deep root soil layers. The AWRA-L model advances one time step (in our case, 1 day) through meteorological forcing and soil water balances (Fig. 1a) to give the amount of water in the soil storage layers. Furthermore, assimilation was conducted one-dimensionally; that is, the AWRA-L model states of a given location were modified independently of any modification made to the states of neighbouring grid cells. Spatial correlation of the resulting model output is however largely maintained because of the spatial coherence of the forcing meteorology. The observation model, H, which maps model states into observation space is, in our case, simply a scaling of AWRA-L top-layer storage estimates by the field capacity value, i.e. S0 =S0FC , to give a relative wetness, denoted x0. Note that this calculation of relative wetness pertains to an unsaturated soil column and does not, for example, account for ponding. Further note that any differences in the interpretation between modelled relative wetness and satellite SM estimates are addressed a data preprocessing step (Section 3.2.1) that ensures the observations have consistent units with what AWRA-L models. In the following we provide only a summary of the salient features of our assimilation approach; details of the computing implementation are provided by Perraud et al. (2013).

2750

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

2.2.1. Forecast step: perturbed meteorological forcing For our investigations we perturbed meteorological forcing data to drive the AWRA-L model and generate ensembles of forecast states. Specifically, we generate an ensemble of Ne forecast states, n oNe , with the i-th element calculated as, xi;f t i¼1

xi;f t

  i ¼ F xi;a i ¼ 1; . . . ; Ne t1 ; Ut

ð3Þ

where F represents the AWRA-L model and superscripts f and a indicate the forecast and analysis state from the current (t) and previous (t  1) time steps, respectively. The ensemble of multivariate Ne forcing data, fUit gi¼1 , with moment characteristics guided by independent knowledge of the errors in the meteorological observations (to be described later in this section), were used to generate the ensemble of model states. Ensuring that the set of ensembles are statistically representative of model state space is critically important to optimal filter performance of the EnKF (Reichle et al., 2008; Turner et al., 2008; Crow and van den Berg, 2010). The ensemble spread should be large enough so that the observations to be assimilated have a chance of exerting constraint on the model estimates (relative to the observation error variance); but not too large as to result in over-fitting of the model (i.e. model estimate is equal to the observation). Alternative method of generating ensembles of model states include simulating model error, perturbed initial conditions, and perturbed parameters (as well as combinations of all the above) (Evensen, 2003). For our investigations, we found that the perturbed meteorological forcing approach, along with the additional ensemble manipulations described in Section 2.2.2, gave adequate ensemble spread for the necessary model error statistics to be computed and therefore the observations to have affect. The meteorological data used were derived following Jones et al. (2009) referred to as ‘BAWAP’ data (in reference to the BoM Australian Water Availability Project for which they were originally developed). These daily climate grids continue to be produced by the BoM (http://www.bom.gov/climate/maps/) and are operationally ingested into the AWRA system. The amount of perturbation was guided by the reported error characteristics of the BAWAP products, which we describe here for the most important forcing variables: radiation, temperature and rainfall.  Daily incoming shortwave radiation (Rs) estimates were derived from geostationary observations and radiative transfer theory with estimated errors in the range 0.8–1.5 MJ m2 d1, which translates into errors <20 W m2. For the purposes of this investigation, we used an additive error for the shortwave radiation perturbation of 50 W m2.  Daily average air temperature (Ta) was calculated from BAWAP minimum (TMIN) and maximum (TMAX) air temperature products as Ta ¼ TMIN þ 0:75ðTMAX  TMINÞ. Jones et al. (2007, 2009) used cross validation to derive errors for TMIN and TMAX products that are on average 2 K, which we have assumed here to be the magnitude of the additive error air temperature perturbations.  Daily total rainfall (Pg) derived from the interpolation of daily gauge observations of the 24-h accumulated rainfall to 9 am on the day of interest. Errors on the BAWAP product have a large north–south gradient, with larger errors in northern Australia compared to the south, as well as larger errors in the arid interior of the continent compared to the coastal regions, reflecting a higher density of rain gauges around the more populated areas (Jones et al., 2009; Chappell et al., 2012, 2013). Jones et al. (2009) report a national average relative error of 60%, which is the value used as the multiplicative error in our rainfall perturbation. Note that our choice of multiplicative error on rainfall, as opposed to the additive errors for radiation

and air temperature, avoids perturbations resulting in negative rainfall for low or zero rainfall records. For each time step, an ensemble of meteorological forcing is generated via Monte Carlo sampling of multivariate normal distributions centred on ðRs ; T a ; P g ÞT with the errors above representing the standard deviations. We also impose a correlation structure on the random sampling, specified by

0

1 1 0:7 0:8 B C 1 0:5 A; C ¼ @ 0:7 0:8 0:5 1 following the approach of Reichle et al. (2007). Note that we replace the perturbation on the rainfall with a univariate random sample between ±0.6Pg, keeping track of whether the original perturbation resulted in an increase or decrease in Pg. The amplitude of the correlation between rainfall and the other meteorological variables may be lost in this step, however the ‘‘direction’’ of the correlation is maintained, and this ultimately is more useful for ensemble generation. 2.2.2. Analysis step: sequential state updating Given satellite soil moisture observations, yt, with specified n oNe error variance R, ensembles of AWRA-L forecast states, xi;f , t i¼1

and observation model, H, we compute analysis states for each ensemble member as,

 1 h   i i;f f t f T xi;a yt  H xi;f þ ei ; t ¼ xt þ P t H HP t H þ R t

i ¼ 1; . . . ; Ne; ð4Þ

where ei  Nð0; RÞ. The ensemble of forecast states are used to calculate the error variance–covariance matrices, specifically,

HPft HT ¼

Ne         T 1 X f i;f H xi;f H x H xft ;  H x  t t t Ne  1 i¼1

ð5Þ

which represents the error variance of AWRA-L modelled soil moisture (alternatively this is known as the forecast error variances of the model states transformed into observation space), and

Pft HT ¼

Ne h i    T 1 X f xi;f H xi;f  H xft t  xt t Ne i¼1

ð6Þ

which is the covariance of forecast model states and forecast model observation. Note that in Eqs. 4–6 H is the linearisation of the observation model H. The degree of influence of the observation on the model states is related to the relative magnitudes of observation and model error variance matrices: e.g. as R tends towards zero, the analysis states become dominated by the observation. Alternatively if HP ft HT  R, observations impart very little or no constraint on model estimation. A common problem with ensemble-based assimilation is the collapse of the ensemble spread, i.e. HP ft HT ! 0, due to the reuse of a small number of ensemble members between time steps. In our investigation we employed both the ‘double EnKF’ of Houtekamer and Mitchell (1998) and covariance inflation technique of Anderson and Anderson (1999) to avoid ensemble collapse. For EnKF to work optimally, care must be taken to ensure that the ensembles adequately represent the model error by having sufficient spread (Reichle et al., 2008; Turner et al., 2008; Crow and Van Loon, 2006). Through preliminary experimentation we found that the approach of double EnKF and covariance inflation (with a factor of 1%) resulted in good ensemble spread that remained long after the influence of a non-zero rainfall perturbation.

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

3. Data sets and observation preprocessing 3.1. Satellite observations Soil moisture (SM) products from two satellite remote sensing systems were used in our investigations. The first was based on data from the (passive) Advanced Microwave Scanning Radiometer for the Earth Observing System (AMSR-E) aboard the Aqua polar orbiting satellite; the second is based on the (active) Advanced Scatterometer (ASCAT) aboard the MetOp-A satellite. Our choice of SM products was guided by independent evaluation of the estimates (e.g. Su et al., 2013), and the study of Draper et al. (2012) that showed the two data sources to be complimentary, leading to improved model soil moisture estimates when assimilated jointly. Note that satellite SM products are derived from emissions (in the case of the radiometer) or backscatter (for the scatterometer) from a much shallower uppermost soil layers (<5 cm for past, current and future microwave remote sensing systems) than is modelled as the AWRA-L top soil layer (8 cm, Fig. 1c). An additional data preprocessing step that has not been included in the version of the AWRA-L data assimilation system used in this work is the smoothing of the satellite SM time series, e.g. via exponential filter (Wagner et al., 1999), to dampen the dynamics of the data and better match the modelled moisture dynamics of a thicker soil layer. We will consider observation smoothing in future modifications to the system, however we do not expect the impact to be dramatic as the observed and modelled thickness are not too dissimilar. 3.1.1. AMSR-E soil moisture product The AMSR-E SM data, known as the Vrije Universiteit Amsterdam (VUA)-NASA product, were derived using the Land Parameter Retrieval Model (LPRM) retrieval algorithm of Owe et al. (2008). LPRM was applied to AMSR-E 6.9 GHz (C-band) brightness temperature observations to yield SM retrieval for the top 1–2 cm of soil, depending on the soil wetness (the wetter the soil, the shallower the observation depth). To facilitate the assimilation into the AWRA-L model, the data were resampled (using nearest-neighbour interpolation) from their original 0.25° resolution to the AWRA-L modelling grid of 0.05° resolution for Australia. The daily SM estimates used here are derived from AMSR-E descending passes, based on the findings of other researchers (e.g. De Jeu, 2003; Draper et al., 2009; Su et al., 2013) that SM retrievals from ‘‘nighttime’’ brightness temperatures show better agreement with in situ measurements. The AMSR-E data cover the period 1 July 2002–30 September 2011. 3.1.2. ASCAT soil moisture product The ASCAT SM product was generated by the Technische Universitat Wien (TUW) using the change detection algorithm of Wagner et al. (1999). Specifically the data are the Surface Degree of Saturations (SDS) values ranging from 0 (driest conditions on record) to 100% (wettest conditions on record). Like the AMSR-E product, the data are derived from C-band measurements and therefore correspond to the top 2 cm of surface soil. As with the AMSR-E SM product the ASCAT SDS data were remapped to the AWRA 0.05° resolution from the original 0.125° data from TUW using nearestneighbour resampling. Both ascending and descending passes of SDS data are averaged following the approach of Liu et al. (2011) to generate the ‘‘daily’’ ASCAT SM data used in the investigation. The ASCAT data were from 1 January 2007–31 December 2011. 3.2. Observation ‘bias’ correction and error characterisation 3.2.1. Cumulative distribution function matching A standard data preprocessing step of data assimilation algorithms is the rescaling of the SM observations to remove any

2751

systematic difference between model and observed time series (the so-called bias correction). Rescaling SM values minimises any potentially detrimental influence of the observation on the model as a result of inconsistency in the definition, measurement and modelling of soil moisture (Koster et al., 2009) by transforming the observations into model space. We use cumulative distribution function (CDF) matching to rescale the satellite SM values (Reichle and Koster, 2004). CDF matching ensures that the statistical distribution of both the satellite data and AWRA-L model time series are the same, so that assimilation only adjusts the model for random variations. 3.2.2. Triple collocation The triple collocation (TC) has become a popular technique amongst the remote sensing and modelling communities for generating spatially explicit error estimates for SM data (Scipal et al., 2008; Dorigo et al., 2010; Miralles et al., 2010; Zwieback et al., 2012). The technique allows for the simultaneous pixel-wise estimation of the error structure, and the cross-calibration of a set of at least three datasets under the assumption that the datasets are linearly related and that the errors are uncorrelated. There are several different implementations of TC. Here we have used the approach of Caires and Sterl (2003). This works with complete triplets and uses a complete 3-dimensional functional relationship to estimate the error variances. Fig. 2 displays the TC error estimates for the two satellite SM products over Australia. The technique was applied to the bias corrected SM data to provide AMSR-E and ASCAT error estimates respectively, with AWRA-L top-layer relative wetness used as the reference product. Details on this application of TC are provided in Henderson et al. (2013). The errors displayed in Fig. 2 are expressed in relative wetness units. Note that there are parts of the continent where TC failed to generate error estimates (white patches, e.g. central Western Australia) due to either a lack of complete triplets or weak correlation between SM data sets and model estimates. For these locations, no data were assimilated into AWRA-L. Broadly we observe that the ASCAT product has lower errors (compared to AMSR-E) across the northern-most parts of Australia and along the eastern seaboard. The AMSR-E product has lower errors in the interior of Australia. These patterns are consistent with the work of Dorigo et al. (2010) for these two SM data sets. Zooming into the region of south east of Australia around the Murrumbidgee catchment (Fig. 2c and d) we can see that the errors on both SM products are highest (AMSR-E more than ASCAT) in parts of the landscape with high fractions of tree cover, notably the alpine regions towards the east (Fig. 2e and f). Errors generally decrease westward into the flat, grazing and cropping parts of the catchment, where ASMR-E errors are on average smaller than ASCAT. Fig. 2c and d also show the location of the OzNet Murrumbidgee network of moisture monitoring probes to be discussed in the next section. 3.3. Surface soil moisture measurements Two sets of surface observations are used in the verification our modelling results. The first are in situ measurements from the long-term, well-studied, moisture-monitoring network in Murrumbidgee catchment in New South Wales, Australia (Smith et al., 2012), known locally as the OzNet network (see Fig. 2c and d); the second are from the newly-formed network of cosmic ray sensors across Australia, known as CosmOz (http://cosmos.hwr.arizona.edu/ Probes/australia.php). Our evaluations focus primarily on the OzNet network because of the longer records of observations, however we also explored the potential of the CosmOz data for model evaluation. The OzNet soil moisture monitoring network is comprised largely of Campbell Scientific water content reflectometry probes

2752

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

(a) ASCAT

(c) M5

M7

M4

M6

ASCAT M3 M2

Yanco Kyeamba

Adelong M1

Error in satellite relative wetness 0.20

0.15 0.10

(b) AMSR-E

(d)

0.05

Fraction

(e) Tree cover (f)

1. 0 0.8 0.6 0.4 0.2 0.0

Fig. 2. Triple collocation estimates of error in satellite soil moisture retrievals in relative wetness units (0–1). Error were derived after both (a) ASCAT and (b) AMSR-E soil moisture time series were CDF matched adjusted to AWRA-L top-layer relative wetness. Note that white areas indicate where the triple collocation algorithm failed to obtain an error estimate. Subfigures (c) and (d) zoom into the region around the OzNet Murrumbidgee network of soil moisture probes, identified as blue squares. Subfigure (e) and (f) show the fraction of tree cover for each AWRA-L model cell, derived from Donohue et al. (2009).

measuring volumetric SM at various soil depths at 63 locations across the Murrumbidgee catchment area (Fig. 2c and d). Specifically, we used 0–8 (or 5) cm measurements to evaluate the model top-layer soil moisture, and the 0–30 cm and 0–90 cm measurements available for 38 of the OzNet sites to evaluate the modelled shallow root-zone soil moisture. For each OzNet probe, we computed daily soil moisture by taking the average of all the measurements (available at 20 or 30 min intervals) within the 24-h period to 9 am local time to be consistent with the BAWAP rainfall forcing data. We assume these daily measurements to be representative of the moisture field for the coincident AWRA-L model cells; where a cell contained two or more probes, we used the arithmetic mean of the in situ measurements for that cell. In total, there are in situ data for 45 sites to evaluate top-layer soil moisture estimates and 37 sites for the shallow root-zone estimates. At the time of writing this paper, we had access to the publicly available OzNet data which largely end 31 May 2011, with the exception of the new set of sensors (labelled Y_A and Y_B) around the Yanco township in New South Wales which end 30 September 2011. The CosmOz network was the first national network of cosmicray SM probes to be established in Australia and is part of a growing international network (http://cosmos.hwr.arizona.edu). The probes are above-ground (standing 1–2 m) passive soil moisture detectors that measure fast cosmic-ray neutrons generated within the soil and air above it. The intensity of these neutrons is moderated predominantly by hydrogen atoms within soil water (Zreda

et al., 2012). The signal source depth of the cosmic-ray probes is influenced by water content itself: higher the moisture in the profile the shallower the effective depth and vice versa. One of the major advantages of the cosmic-ray probe over traditional measurement approaches is its measurement scale which is approximated by circle with a diameter of 600 m, i.e. 30 ha (Desilets and Zreda, 2013). Cosmic-ray probes have been demonstrated to provide accurate estimates of root-zone soil moisture content for depths between 15 and 70 cm (Franz et al., 2012a,b). The CosmOz network is comprised of 10 probes situated at a variety of locations across Australia. Full details of data processing and probe calibration for the CosmOz network are provided by Hawdon et al. (2014). At the time of writing this paper, 7 probes were collecting data for some period overlapping with the analysis period. The site location and characteristics of the CosmOz probes used in our study is provided in Table 1. The data are logged at 1-h time intervals, however we used a 24-h average to represent the daily moisture measurement. The two SM monitoring networks coincide at Yanco, NSW. Fig. 3 is a simple time series plot of daily measurements from the OzNet sensors and CosmOz probe near Yanco. The shaded band of values represent the minimum and maximum from a set of OzNet sensors (labelled Y_A and Y_B) within an approximate 2-km radius of the CosmOz probe, whose measurements are displayed as dots. The solid line is the mean of the OzNet data, while the dashed line indicate the measurements from a single sensor (Y_10) closest to the cosmic ray probe. The data show high level of agreement between

2753

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762 Table 1 Site and location information of the Australian cosmic-ray soil moisture monitoring probes (CosmOz) used in the evaluation of AWRA-L root-zone soil moisture. Station Name

Latitude/Longitude

Records start

Elevation (m)

Measurement depth range (cm)

Vegetation cover/land use

Australian soil classification

Baldry, NSW Daly, NT Robson Creek, Qld Tullochgorum, Tas. Tumbarumba, NSW Weany Creek, Qld Yanco, NSW

32.8710/148.5380 14.1592/131.3881 17.1200/145.6300 41.6694/147.9117 35.6560/148.1520 19.8820/146.5360 35.0050/146.2992

30 May 2011 7 June 2011 28 October 2010 15 December 2010 3 April 2011 1 December 2010 1 April 2011

438 75 715 285 1200 287 124

6–33 15–46 7–23 6–31 6–14 9–32 9–26

Open grassland/Grazing Tropical savannah/Grazing Tropical rainforest/National park Improved pasture/Grazing Wet eucalypt forest/State forest Open woodland/Grazing Open grassland/Grazing

Chromosol Kandasol Dermosol Sodosol Kandosol Chromosol Sodosol

All AWRA-L simulations ran unconstrained, driven only with perturbed forcing, from 1 January 2000 to 31 December 2005. Six years of model spin-up was chosen after preliminary experimentation revealed that it was of sufficient length for some of the water balance terms (notably the river water stores) and the ensemblederived error statistics to be free of initial conditions. Assimilation commences 1 January 2006 and ends 31 December 2011.

OzNet

0.4 0.2

4.1. Data assimilation experiments

0.0

Soil moisture [m 3 / m-3 ]

0.6

CosmOz

Jan

Mar

May

Jul

Sep

Nov

Jan

Date [2011] Fig. 3. Time series of soil moisture measurements (m3 m3) around the Yanco town site, NSW, from OzNet sensors (lines) and CosmOz probe (dots). OzNet data correspond to the integrated 0–5 cm soil moisture measurement at a point, while CosmOz data correspond to integrated soil moisture measurement for depths of 0– 20 cm or 0–40 cm, depending on the degree of soil wetness, and 30 ha area around the probe. The shaded band represents the range of measurements (solid line is the mean) from OzNet sensors within a 2-km radius of the CosmOz probe, while the closest sensor (Y_10) to the probe is the dashed line.

the two data sources, despite the OzNet data being 0–5 cm integrated SM measurement at points and the CosmOz probe measuring a depth of emissions from the top 20 to 40 cm of soil (depending on soil wetness) over a 30 ha area. This may not be too surprising given that a subset of the OzNet sensors are used in the calibration of the CosmOz probe. Nevertheless the agreement illustrates the potential of the new data for model evaluation and complements the traditional sensor systems for ground-based moisture monitoring. 4. Results Here we evaluate the results of assimilating the AMSR-E and ASCAT SM data into AWRA-L against the OzNet and CosmOz in situ SM data. The CDF matched satellite SM data were used along with corresponding TC error estimates for both products. AWRA-L simulations were executed cell-wise (independent cells) across Australia at 0.05° intervals using CSIRO’s parallel computing infrastructure and large volumes (i.e. >1 Tb) of disk storage to accommodate the ensembles of model outputs. Details of this implementation are given in Perraud et al. (2013). This was not essential for the purposes of the experiments, however were part of a larger continental scale evaluation of AWRA-L’s performance for other parts of the water balance prior to operational implementation within the BoM.

4.1.1. Simulations Four experiments were conducted: assimilation of AMSR-E alone; assimilation of ASCAT alone; joint assimilation of AMSR-E and ASCAT (i.e. simultaneously); and open-loop (no assimilation) to serve as a reference. Given ASCAT SM time series begins on 1 January 2007, we evaluated the results of all experiments over the common period 1 July 2007 (6 months into the ASCAT time series) – 31 May 2011 (end of OzNet data). A comparison of open-loop AWRA-L simulations for top-layer relative wetness (using the ensemble mean as the representative SM estimate), AMSR-E and ASCAT moisture data (CDF matched) at the OzNet sites is given in Fig. 4. Generally both AWRA-L open-loop simulations and the satellite SM products are highly correlated with the 0–8 cm in situ data (or 0–5 cm for the Yanco sites Y_A and Y_B), with average correlation of 0.72 (range between 0.6 and 0.9). For some of the Y_A and Y_B sites correlation of AWRA_L simulations and AMSR-E data drops to about 0.5. For comparison the correlations of the original satellite products (i.e. without CDF matching) are displayed in Fig. 4 as grey squares. Overall the CDF matching made little difference to the correlation of the satellite products across the OzNet sites. There were some sites, however, around the Yanco area where the CDF matching of the AMSR-E time series slightly improved correlation of the satellite products. Also the CDF matching resulted in a slight reduction in correlation of both satellite time series at the Y_A sites. Across the 45 OzNet sites AWRA-L estimates outperform (i.e. have higher correlation) the satellite products at 21 sites compared with ASCAT, and 28 sites compared with AMSR-E. There was, however, broad consistency in the correlation ‘‘pattern’’ across the OzNet sites between modelled and satellite SM, i.e. agreement in the ‘‘direction’’ of the high (e.g. K_11 with r = 0.9) and low (e.g. M_3 with r = 0.6) correlations. Also displayed in Fig. 4 are the site-wise triple collocation (TC) errors for both satellite products (in relative wetness units). The AMSR-E product had higher errors for the more vegetated area around M_1, M_2, Adelong, and Kyeamba compared to ASCAT; while the ASCAT had higher errors around the Yanco sites. Both products had on average an error of 0.09 relative wetness units, which is consistent with ‘‘approximate’’ 10% relative error associated with these products (e.g. Draper et al., 2012). For each experiment, 100 member ensembles of AWRA-L soil water storage estimates were generated at each time step. Ensembles are summarised for output by their 0th, 25th, 50th (median),

2754

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

0.8

0.20

0.6

Correlation

0.4

0.10

0.2

Original

0.05

AWRA-L

open loop

CDF matched

TC Error

0.0

0.00

0.8

0.20

0.6

0.15

0.4

TC error (relative wetness units)

0.15

0.10

0.2

Original

open loop

0.05

CDF matched

TC Error

0.00 A_5

A_3

A_1

K_13

K_8

K_11

K_6

K_4

K_1

Y_10

Y_B3

Y_A9

Y_13

Y_A3

Y_8

Y_11

Y_6

Y_4

Y_2

M_7

M_5

M_3

M_1

0.0

Fig. 4. Correlation coefficient for open-loop AWRA-L top-layer simulations and the satellite soil moisture products evaluated against in situ measurements at each OzNet sites (x-axis) over the period 1 July 2007–31 May 2011. Correlation of AWRA-L (open-loop) estimates are black hollow circles and the CDF-matched ASCAT (top panel) and AMSR-E (bottom panel) correlations are grey hollow squares, with the grey solid squares representing the original (i.e. no CDF matching) satellite products. The triple collocation (TC) errors for the CDF-matched satellite products are also presented as blue triangles. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 5. AWRA-L top-soil layer relative wetness (w0) estimates across Australia for 7 July 2009: (a) analysis estimate after the assimilation of ASCAT soil moisture into AWRA-L; and (b) the analysis increment (difference between analysis and forecast estimates) highlighting the location and amount of influence of the satellite data. Subfigure (c) shows a time series of AWRA-L simulations compared to satellite SM estimates (white dots) for a Yanco OzNet site, Y_3. Ensemble median and interquartile range for open-loop simulations highlighted as a black line and grey shaded area, respectively. Analysis estimates after the assimilation of ASCAT and AMSR-E data in top and bottom panels respectively, with ensemble median and range as solid blue line and cyan shading.

75th 100th percentiles and ensemble mean. To illustrate, Fig. 5a depicts the ensemble median of AWRA-L analysis top-layer relative wetness for the whole of Australia for 7 July 2009 resulting from the assimilation of ASCAT SM. The difference between analysis

and forecast SM estimates (the analysis increment, Fig. 5b), clearly show the influence of the satellite data on the model estimation. The assimilation of ASCAT, in this example, resulted in moisture being added (blue colouring) to the model for parts of Western

2755

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

Australia and southern Victoria, and moisture being subtracted from some other area (brown colouring). Time series of AWRA-L top-layer relative wetness estimates, with and without assimilation are provided in Fig. 5c. The model output is displayed for 1 April 2010–30 April 2011 for the Yanco Y_3 OzNet site. Here, the open-loop and analysis relative wetness estimates are displayed in black and blue, respectively, with the shaded region representing the ensemble interquartile range (i.e. difference between 75th and 25th percentiles). For comparison, the CDF adjusted ASCAT and AMSR-E SM data for the sites are displayed as white dots in the top and bottom panels respectively. The AWRA-L open-loop simulations can be seen to match the peaks in SM detected by both satellite data sets. However the model tended to dry out the top-soil layer quicker than the observations suggest, and is particularly noticeable for the October 2010–January 2011 period. Assimilating ASCAT and AMSR-E data into AWRA-L brought the analysis estimates into closer agreement with the satellite SM values. Also note that the saturation in the AMSR-E time series around November 2010 was due to the CDF matching applied. For this site, between 16 October–30 November 2010, satellite SM values were retrieved that were greater than the 99th percentile over whole AMSR-E time series (ie. 0.4 m3 m3). The result points to the need for stricter quality control measures to be applied to the satellite time series data in future developments of the AWRA DA system. 4.1.2. Model evaluation Quantitative evaluation of the AWRA-L assimilation results were based on spatially and temporally aggregated OzNet data (Section 3.3). The daily CosmOz data were used only to evaluate root-zone moisture estimation. The number of OzNet data available for the comparison varies between sites: the older network (M_1 – M_7, Y_1 – Y_13, K_1 – K_14, and A_1 – A_5) has between 670 and 1430 daily data points over our evaluation period; the newer Y_A and Y_B probes around Yanco have around 400–500 (two sites having fewer than 100 points). Note that due to the incongruence of what AWRA-L models (water storage or relative wetness) and what the OzNet probes measure (volumetric SM), the main metric of comparison here is the correlation of point time series, acknowledging that converting AWRA-L output into volumetric units for root mean squared error or bias calculations requires linear scaling of the model output using the soil map hydraulic property information (Fig. 1) which may introduce bias.

significant difference (at p-value = 0.05) in ra  r0 with or without the assimilation of satellite SM. However, for the OzNet sites where satellite data were more highly correlated with in situ measurements than open-loop simulations (Fig. 4) a general improvement in ra, particularly for sites M_6, Y_6, Y_11, Y_13, Y_A1, Y_A3, and Y_B1, was observed. Note that these same 7 sites have statistically significant increases (i.e. the differences ra  r0 were outside the 95% confidence limits p-value = 0.05) in ra for the three assimilation experiments. There were some sites where the assimilation of satellite product resulted in ra < r0, despite the satellite SM product having higher correlation than the open-loop estimates: for AMSR-E they were sites M_4 and K_5; for ASCAT they were Y_A9, K_1, K_2, and K_5. For these sites, the degree of degradation was lessened through the joint assimilation of AMSR-E and ASCAT. This was most evident for site Y_A9, where the assimilation of ASCAT led to a 5% reduction in correlation compared to open-loop simulations, however the joint AMSR-E and ASCAT assimilation in AWRA-L resulted in an a 5% increase in correlation (i.e. from ra = 0.66 to ra = 0.73). Note that at site Y_A9 the TC error for ASCAT was higher (0.09 in relative wetness units) than AMSR-E (0.05), despite the higher correlation of ASCAT with the in situ measurements. There were some sites where ra > r0 despite the open-loop simulations having higher correlation than the satellite retrievals. For example, the assimilation of AMSR-E resulted in statistically significant increases from r0 = 0.85 to ra = 0.88 and from r0 = 0.83 to ra = 0.88 for sites Y_1 and Y_4, respectively. For these sites, the

30

(a)

AMSR-E

(b)

ASCAT

(c)

AMSRE + ASCAT

20 10

0 -10 30 20 10

4.2. Top-layer soil moisture analysis 0

The correlation between the ensemble mean top-layer SM estimates, x0, and in situ 0–8 cm data for each of the 45 OzNet sites was computed for the period 1 July 2007–31 May 2011. Fig. 6 shows plots of correlation for the open-loop estimates (r0) along the x-axis for each OzNet site and the percentage relative difference between the analysis estimates (ra) and open-loop after the assimilation of (a) AMSR-E, (b) ASCAT and (c) both data sets jointly. Highlighted in grey are those sites from Fig. 4 where the satellite product had correlation greater than open-loop x0, and in white those sites where the correlation of AWRA-L estimates was greater. Also plotted in Fig. 6 are the confidence limits on the difference in correlation derived via Fisher Z-transformations for a 95% significance level for a sample size of 900 points (the average number of points across OzNet sites). Points falling above the x-axis in Fig. 6 indicate that the assimilation of SM products improved AWRA-L top-layer soil moisture estimates at a given OzNet site (i.e. ra > r0), while points below the axis indicate degradation due to the assimilation (ra < r0). For most OzNet sites, the results suggest that there was no statistically

-10 30 20 10 0 -10 0.2

0.4

0.6

0.8

r0 Fig. 6. Percentage relative difference in correlation between AWRA-L simulations of top-layer relative wetness (w0) and OzNet measurements before (r0) and after (ra) assimilation of satellite soil moisture products. Grey dots correspond to OzNet sites where the correlations of satellite soil moisture product was greater than open-loop w0 and white dots the sites where the correlation of w0 was greater than satellite data (from Fig. 4).

2756

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

correlation of AMSR-E products with in situ data was only marginally smaller than r0, and the TC error is also quite low (0.05).

limits of what would be generally considered the root-zone. However there are some sites (e.g. K_12 and Y_7) where the addition of Sd for any proportion results in a reduction of the correlation. Here we used the ensemble means to evaluate the AWRA-L rootzone moisture SM estimates against the in situ measurements at 37 OzNet and the 7 CosmOz sites. As above, we focus on the correlation of the time modelled and in situ time series as the performance metric.

4.3. Root-zone moisture analysis

Correlation with 0-90 cm measurement

Root-zone moisture estimates from AWRA-L were constructed by summing modelled top-layer, shallow- and deep-root soil layers in proportions that, based on the estimated soil layer thicknesses (Fig. 1), were equivalent to the measurement depths (0–30 cm and 0–90 cm for OzNet, various depths for CosmOz (Table 1). The modelling of the 0–30 cm root-zone is achieved by adding S0 and Ss, while the 0–90 cm root-zone requires the addition of a proportion of Sd. Fig. 7 illustrates the effect of adding different proportion of Sd to the root-zone (open-loop) on correlation against 0–90 cm in situ measurements for some OzNet sites. For most site the addition of Sd increase correlation to a point, and then the correlation reduces with addition of more Sd. Encouragingly, the ‘‘peak’’ in correlation coincides generally with the inferred AWRA-L 0–90 cm root-zone thickness, identified in Fig. 7 by dots, with the exception of a minority of sites (e.g. Y_6) where correlation grew beyond the

4.3.1. Shallow-root zone and OzNet As with the top-layer evaluations, improvements (or degradations) were assessed at each site by comparing analysis and open-loop correlations, ra and r0 respectively. The percentage relative difference in ra and r0 for the OzNet sites at the 0–30 cm and 0–90 cm depths is displayed in Fig. 8. As in Fig. 6, points falling above the x-axis indicate an improvement in correlation as a result of the assimilation of satellite SM products (points below indicate a degradation), and the 95% significance limits are displayed in Fig. 8 as dotted lines. We can see in Fig. 8 that, for OzNet, the assimilation of satellite SM into AWRA-L generally improves AWRA-L estimates for the 0– 30 cm layer at most of the sites, with statistically significant improvements for 40% and 43% of sites for AMSR-E and ASCAT assimilation respectively. The assimilation of AMSR-E resulted in a decrease in ra relative to r0 for the 0–30 cm model estimates for 4 sites: M_7 (by 1%), Y_13 (6%), K_8 (3%) and K_14 (2%) (Fig. 8a). The assimilation of satellite SM shows the greatest improvement (r0 = 0.33 and ra = 0.45) in model estimation for the OzNet site Y_6, which also happens to be a site where the assimilation of satellite SM products improved top-layer SM estimation by more than 10% (Fig. 6). The joint assimilation of AMSR-E and ASCAT resulted in improvements in 0–30 cm root-zone moisture estimation for almost all the sites, with only Y_13 (ra < r0 by 3%) and M_7 (ra < r0 by 1%) showing degraded estimates relative to open-loop (Fig. 8c). AWRA-L 0–90 cm root-zone moisture estimates show significant improvements in correlation (increases of >30%) after the assimilation of satellite SM for sites where r0 < 0.6 (Fig. 8d–f). The most dramatic improvement, with increases in correlation of >85%, is seen for sites Y_1 (where r0 = 0.32) and Y_6 (where r0 = 0.33), where the assimilation of ASCAT (for example) increased correlation from

0.7 K_1

0.6

M_3, A_3

0.5

Y_6

0.4 0.3 K_12

0.2 Y_7

0.1 0 0%

5%

10%

15%

20%

Percentage of Sd contributing to modelled root-zone Fig. 7. Correlation of AWRA_L root-zone estimates with 0–90 cm measurements for a subset of OzNet sites. The lines illustrate the change in correlation resulting from the use of different proportions of deep root soil storage (Sd) in the calculation of the root-zone moisture.

AMSR-E 40

(a)

(b)

ASCAT

AMSRE + ASCAT (c)

(d)

(e)

(f)

20

100 x

ra − r0 /r0

0 -20 -40 100 80 40 20 0 -20 -70 -90

0.2

0.4

0.6

0.8

0.4

0.6

0.8

0.4

0.6

0.8

r0 Fig. 8. Percentage relative difference in correlation between AWRA-L shallow root-zone soil moisture estimates and OzNet measurements for 0–30 cm (top panel) and 0–90 cm (bottom panel) soil depths before (r0) and after (ra) the assimilation of satellite soil moisture products.

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

0.3 (r0) to 0.6 (ra). Note that for both Y_1 and Y_6, the assimilation of satellite SM improved AWRA-L top-layer SM estimation (Fig. 6) and also the addition of Sd to the modelled root-zone greatly increased the open-loop correlation with 0–90 cm measurements (e.g. Fig. 7). For site Y_7, however, the assimilation of either satellite SM product reduced the correlation of AWRA-L top-layer estimation and, furthermore, the addition of Sd degraded the open-loop correlation with 0–90 cm measurements. Not surprisingly, therefore, the assimilation of satellite data also resulted in a >70% reduction in correlation with 0–90 cm measurement when compared with AWRA-L open-loop at site Y_7 (point to the lower left corner of Fig. 8d–f). For the majority of OzNet sites (>60% of them), the joint assimilation of satellite products resulted in improvement or no statistically significant improvement (at p-value 0.05 significance level) in 0–90 cm SM estimation compared to r0. Site K_12 shows the value of joint assimilation of both satellite products on AWRA-L 0–90 cm root-zone SM estimation. The assimilation of AMSR-E resulted in degradation of correlation of 3% relative to the open-loop (r0 = 0.37, Fig. 8d). However the joint assimilation of AMSR-E and ASCAT resulted in a 30% improvement in correlation (ra = 0.47, Fig. 8f). This is despite there being a decrease in correlation between model and 0–90 cm SM through the addition of Sd. (Fig. 7). We further note that for K_12, top-layer estimation increases correlation with 0–8 cm measurements from r0 = 0.65 to ra = 0.72 after the assimilation of ASCAT whereas for AMSR-E the result was a slight decrease in correlation (by 1%). Overall, the assimilation of AMSR-E and ASCAT soil moisture estimates had a consistent positive impact on AWRA-L 0–30 cm SM estimation, where the improvements in performance are most apparent (Fig. 8a–c). This suggests that the assimilation of satellite SM products had the most affect on the AWRA-L model shallow root soil layer, Ss. Fig. 9 displays the cumulative distributions of the analysis increments of the soil water storage states (i.e. the difference between analysis and forecast estimates of the model state variables) normalised by the forecast estimates, i.e. ðxa  xf Þ=xf . This statistic was pooled across all OzNet sites for only those times when there were satellite data to be assimilated (i.e. non-zero analysis increments) to avoid a dominance of zeros influencing the distributions. In terms of variances, Fig. 9 shows (in deceasing order) the largest variance (i.e. the greatest analysis increment on average) corresponds to the shallow-root layer, followed by toplayer, and finally the deep-root layer storages.

4.3.2. Shallow root-zone and CosmOz AWRA-L shallow root-zone moisture estimation was evaluated against daily average measurements from 7 CosmOz probe sites across Australia (Section 3.3, Table 1). The evaluation period in this case was from the start of the CosmOz data archive (28 October 1.0

Probability

0.8

Top-layer Shallow Deep

0.6 0.4 0.2 0.0 -10

-5

0

5

10

(xa − xf )/xf Fig. 9. Cumulative distribution of the analysis increments of the AWRA-L soil water storage states (normalised by the forecast states estimates) pooled across the OzNet site and only for those times when satellite SM were available for assimilation.

2757

2010 with Robson Creek site) to end of meteorological forcing data availability (30 June 2012). Our first set of evaluations against CosmOz data, required scaling both measured and modelled quantities to consistent set of unit. We summed AWRA-L open-loop top- and shallow-root layer water storages to represent the modelled root-zone and normalised the results to have zero temporal mean and unit variance for the evaluation period, generating normalised wetness time series. These are displayed for the CozmOz sites, along with similarly treated probe measurements, in Fig. 10. Note that normalising the data in this way avoids the need to use soil hydraulic parameters, and keeps the focus on correlation of the two data sets. In Fig. 10 the interquartile range (IQR) of AWRA-L ensemble estimates is represented by the grey bands, while the cosmic-ray probe data are blue dots. The AWRA-L and CosmOz SM estimates show great similarity in the soil wetness pattern, with peak and troughs generally coinciding. The correlation between AWRA-L estimates and CosmOz data was generally high (Table 2). The correlation of AWRA-L shallow root-zone moisture estimates against the CosmOz probes over the period 1 July 2007–31 December 2011 is displayed in Table 2. The assimilation of satellite SM slightly improved AWRA-L shallow root-zone estimation for Baldry and Yanco CosmOz sites, but had little or no affect for the other sites. At the Baldry site, the assimilation of the AMSR-E SM degraded results relative to the AWRA-L open-loop estimates. However, at the same site, the assimilation of the ASCAT product improved agreement between model and probe data compared to the open-loop estimates. AMSR-E and ASCAT products had similar TC errors (in relative wetness units, Table 2) and therefore in the joint assimilation, both products contributed equally to the analysis. Whilst the result of the joint assimilation was a compromise in performance, the correlation was closer to the ASCAT-only result. Similar compromise in performance through the joint assimilation was observed for the Yanco site, however single data assimilation was an improvement compared to open-loop for both SM products. The TC errors (Table 2) provide a partial explanation as to why the assimilation of satellite SM resulted in no or little difference in AWRA-L performance for the majority of CosmOz sites, with the exception of Tumbarumba which may be a model performance issue (need further investigation to understand why). AMSR-E TC errors are generally high (>0.14) for these sites, which can be attributed to the well-known poorer quality of the passive microwave SM retrievals in areas of dense woody vegetation cover (e.g. rainforest, Robson Creek) and areas proximal to the coast (e.g. Tullochgorumn) and water bodies. In contrast ASCAT TC errors are low (0.07–0.09). And while ASCAT assimilation was observed to improve AWRA-L performance at the Tullochgorum site (perhaps due to the finer resolution of the ASCAT retrievals compared to AMSR-E), the general null effect suggests the model ensemble spread may be too narrow (e.g. IQR of 0.04 relative units on average) to allow the influence of the observation. Finally, the CosmOz probe data were used to evaluate AWRA-L shallow root-zone estimates for various thicknesses, varying the root-zone by 5 cm increments between 0–5 cm and 0–2 m. The estimates were constructed in the same way as described at the start of Section 4.3, namely by weighted sums of the storages in the top-, shallow-root and deep-root layers (using the estimates layer thicknesses, Fig. 1) to get to the desired root-zone thickness. The correlation between AWRA-L model estimates and probe measurements for each site was used as the measure of agreement. The correlation rapidly increased from 0–5 cm to 0–20 cm thickness, slowly began to plateau over a broad range of thicknesses, reached a maximum, and then decreased for thicknesses beyond 1.3 m. Table 3 shows the range of shallow root-zone thicknesses centred on the maximum (rmax) and varying by ±0.02 correlation units. The

2758

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

Normalised wetness

4 3 2 1 0 -1

Normalised wetness

-2 4 3 2 1 0 -1

Normalised wetness

-2 4 3 2 1 0 -1

Normalised wetness

-2 4

2011

2012

3

CosmOz

2

AWRA-L (ensemble interquartile range)

1 0 -1 -2 2011

2012

Fig. 10. Time series of AWRA-L shallow root-zone soil moisture estimates and cosmic-ray probe moisture measurements (dots) at the CosmOz sites (time series normalised for visual comparison to zero mean and unit variance for 28 October 2010–30 June 2012). Note that the AWRA-L estimates displayed correspond to open-loop simulations, with the grey band representing the interquartile range of ensemble spread.

Table 2 Evaluation of AWRA-L shallow root-zone moisture estimates (i.e. sum of top- and shallow-root layer water storages) against measurements from the CosmOz probes: N is the number of daily measurements for each CosmOz probe between 1 July 2007–31 December 2011; triple collocation (TC) errors in relative wetness units for each satellite product are displayed; and open-loop (r0) and analysis (ra) correlation for the AWRA-L estimates against probe data are displayed. Baldry

Daly

Robson

Tullochgorum

Tumbarumba

Weany

Yanco

398 0.10 0.12

390 0.14 0.07

611 0.15 0.09

559 0.15 0.08

349 0.11 0.10

577 0.16 0.08

451 0.08 0.08

Profile moisture: S0 + Ss (Evaluation 1 July 2007 – End of satellite time series) r0 0.79 0.94 0.86 ra (AMSR-E) 0.76 0.94 0.87 a r (ASCAT) 0.83 0.92 0.86

0.82 0.80 0.84

0.51 0.52 0.51

0.94 0.93 0.94

0.78 0.79 0.83

ra (AMSR-E and ASCAT)

0.82

0.52

0.94

0.80

N TC error (AMSR-E) TC error (ASCAT)

0.82

0.94

0.86

Table 3 Range of AWRA-L modelled shallow root-zone layer thickness (mm) exhibiting highest correlation (rmax) with cosmic ray probe probe data for each CosmOz site.

Range (mm) rmax Shallow root-zone layer thickness (mm) at rmax ± 0.02

Baldry

Daly

Robson

Tullochgorum

Tumbarumba

Weany

Yanco

0.79 200–750

0.93 100–400

0.88 250–900

0.90 350–900

0.49 300–1250

0.94 200–1000

0.73 200–1200

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

results demonstrate the greatest agreement between model estimates and CosmOz data were for root-zone thicknesses generally regarded by most land surface models as the ‘‘root-zone’’. Moreover the thicknesses are consistent with the reported effective depths of emissions for the cosmic-ray probes (Table 1). 5. Discussion 5.1. AWRA soil moisture data assimilation framework The AWRA system provides comprehensive water balance estimates across Australia at 0.05° resolution and at daily time steps. The goal of the system is reconstruct water balance at an unprecedented spatial scale from the start of the 20th century to present, and place the previous year’s water balance in the context of patterns and trends of water resources over the last one hundred years. The data assimilation framework embedded in the AWRA system was developed to work in conjunction with the AWRA calibration framework (Viney et al., 2011) and constrain model estimates with combinations of ground-based and remotely-sensed data, resulting in water balance estimates of unparalleled accuracy and internal consistency amongst variables not realised in Australia to-date. The focus of the current study was on the evaluation of the soil moisture data assimilation component of the AWRA system. Currently, it is only the soil water storage components of the system that are set up to routinely assimilate satellite soil moisture for the entire country (coupling to river and groundwater components is not yet developed). The assimilation has considered the AMSR-E and ASCAT time series of soil moisture products, however it can be readily extended to ingest soil moisture products from older passive/active microwave systems (e.g. TMI, SSMI, ERS, ASAR) and the suit of new (SMOS, AMSR2) and planed (SMAP) dedicated satellite soil moisture missions. The complementary active–passive soil moisture sensing approach is well-known and was observed in this work in the case where the assimilation results for one satellite product was poorer, the joint assimilation with the other product improved results. The evaluations presented have focussed on the AWRA-L model estimates of the root-zone moisture. The primary relevance of the root-zone moisture estimates from AWRA system is for drought monitoring, where the data can provide indicative trends, deficits and surpluses of moisture relative to a specified climatology. The data may also serve as independent verification for other models, particularly the land surface components of numerical weather prediction systems. Currently in Australia the BAWAP system (Raupach et al., 2009; King et al., 2009) generate the closest equivalent to the root-zone moisture grids from AWRA-L, however the model output are unconstrained by satellite observations, nor are the model components as extensively calibrated as in the AWRA system. Moreover the AWRA system is unique as groundwater and river processes are also modelled, thereby closing the water balance. 5.2. Assimilation method: the ensemble Kalman filter The ensemble Kalman filter (EnKF) was chosen as the method for sequentially updating AWRA-L model soil water storages (the prognostic states of the model) when and where the satellite data were available. One of the main motivations in implementing the EnKF as part of the AWRA system was the flexibility of the method to changes in the dynamical model which, in the case of AWRA-L, may undergo periodic development and refinement as user requirements change, necessitating regular re-calibration as additional metered data or new landscape parameter sets become available. Furthermore, unlike variational data assimilation

2759

counterparts typically used in operational meteorological centres for weather and climate predictions, the EnKF does not require tangent linear of adjoint coding. Perturbed meteorological forcing drives the ensemble generation for the EnKF in the AWRA system. The perturbations were guided by published error statistics in the forcing meteorology (obtained from the BAWAP system, Jones et al., 2009). These errors are often quoted in terms of national average values and for simplicity we have assumed that the single values are representative of the error for the whole country. This assumption is particularly invalid for the BAWAP rainfall products as Renzullo et al. (2011), for example, observed in the generation of their own daily rainfall analyses for Australia. The skewed spatial distribution of the underlying rain gauge network towards the more populated parts of the country result in greater modelling errors relative to the temporal dynamics in the arid inland regions and lower errors towards the coastal fringe. In future the magnitude of the rainfall perturbation (as well as the air temperature and shortwave radiation data) will be made to spatially vary following the error distribution derived from the works of Chappell et al. (2012, 2013). An ensemble of AWRA-L model forecast and analysis states (water storages and fluxes) are generated and stored as grids on local computing infrastructure. The ensemble of model estimates may be used to represent estimation uncertainty in water balance terms. However, current implementation of the EnKF does not explicitly represent model error. Future work will explore the implementation of stochastically modelled AWRA-L model error, following the examples of Reichle et al. (2002) and Crow and van den Berg (2010) which will result in better representation of AWRA-L uncertainty and aid in optimal implementation of the EnKF. 5.3. Top-layer soil moisture estimation Grids of AWRA-L top-layer soil moisture (SM) estimates were evaluated across the Murrumbidgee network of in situ (OzNet) moisture sensors. Top-layer SM estimation was improved through the assimilation of satellite SM products (AMSR-E and ASCAT) only for those sites where the satellite data had higher correlation with the ground-based data than AWRA-L open-loop estimates. For these sites, improvements in correlation of as high as 30% were achieved through the assimilation of satellite SM. The degree of improvement was observed to decrease as the correlation of AWRA-L open-loop increased. There were some OzNet sites where the assimilation of AMSR-E data led to degradation in the correlation of modelled estimates, despite the correlation of the satellite data alone being higher than AWRA-L open-loop estimates. These, however, were not statistically significant. The result highlights the challenge for data assimilation to improve model estimates when the model performance is quite good. The correlation statistics of AWRA-L open-loop estimates across the OzNet network were already quite high, with value typically above 0.6 (0.75 on average). Generally the improvements, or degradations, in AWRA-L top-layer SM estimation resulting from the assimilation of satellite SM were generally not statistically significant. When model performance is good, the adjustment by satellite data is not needed, and indeed we have seen that performance was not enhanced significantly through assimilation. However, when model performance is poor, and provided the satellite data are more highly correlated with in situ data than the model (and that satellite SM errors are low), the improvements through assimilation can be substantial. 5.4. Root-zone moisture estimation Grids of AWRA-L shallow root-zone moisture estimates (limited to the top 0–1 m of the soil column) were evaluated against in situ

2760

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

measurements from the OzNet sites as well as the new proximal sensing cosmic-ray probe network (CosmOz) around Australia. A consistent improvement in AWRA-L model performance through the assimilation of satellite SM was on the shallow-root soil layer. Proportionally the analysis increment was observed to have higher variance (i.e. greatest impact) on the shallow-root soil layer, followed by the top-layer and finally the deep-layer (Fig. 9). Recall that these layers are adjusted simultaneously in the analysis step of the EnKF, where they are brought into closer agreement with the satellite SM. The error variances used in the EnKF (Eqs. (5) and (6)), derived through ensemble statistics at each time step, combine to apportion greater weight to the shallow-root soil water state relative to the other storage layers. This is likely due to the strong coupling between the top and shallow-root layers. The assimilation of satellite SM resulted in improvements in AWRA-L estimates for the 0–30 cm at almost all OzNet sites. The model performance at the 0–90 cm layer was more variable, but could be related to performance of the assimilation on the AWRA-L top-layer simulations (Fig. 6), which is indicative of the coupling between top- and deep-soil layers, and the impact of including the deep root soil layer storage (Sd) in the calculation of root-zone moisture (Fig. 7). Generally improvements in AWRA-L 0–90 cm estimation were observed for sites where the assimilation of satellite SM improved model top-layer SM and for which the addition of the deep soil layer storage improved openloop simulations (compared to simply S0 + Ss). For site Y_7, however, the assimilation of satellite SM did not improve top-layer estimation, and because of the negative influence of Sd on openloop estimates of 0–90 cm moisture (i.e. Fig. 7), the assimilation of satellite SM resulted in significantly degraded performance. In contrast, for site K_12, the assimilation of AMSR-E did not improve AWRA-L top-layer performance but the assimilation of ASCAT did. And despite the inclusion of Sd degrading open-loop estimates, the result of the joint assimilation of AMSR-E and ASCAT was an improvement in 0–90 cm estimation. The influence on the deep root soil layer on AWRA-L root-zone estimation (and the impact of SM assimilation) needs to be better understood and remains the topic of further exploration. The increment of the EnKF applied to the deep root soil layer may be too large, which appears to lead to an improvement in root-zone SM estimation for some sites and a degradation for others. Nevertheless, the state updating of the deeper soil layer may need to be dampened to better represent deep soil moisture dynamics. This could, for example, be achieved via the specification of a background/forecast error variance matrix that prescribes correlation structure between soil layers as in Parrens et al. (2014).

5.5. Evaluation data Evaluation of AWRA-L shallow root-zone moisture estimates was conducted using ground data from both a traditional in situ TDR-based sensor network (OzNet) and the emerging new proximal sensing technology based on cosmic-ray sensor technology (CosmOz). Both technologies have their respective pros and cons: e.g. OzNet probes are direct single-point measurements, but very accurate and constant in sensing depth; CosmOz probes are integrated in space (500 m radius of probe) and therefore suited for modelling and (some) satellite verification scales, however are themselves retrievals for temporally-varying soil depths (shallow root-zone). We have used correlation as the primary metric for evaluation as what is ‘‘measured’’ on the ground is often incongruent with what is modelled (i.e. representativeness). Relationships relating processes at the modelling scale with that of the observation (ground-based or remotely-sensed) are required for further quantitative evaluations.

The CosmOz probe data were most highly correlated for AWRA-L root-zone estimates over the range of thickness layers typically identified as the ‘‘root-zone’’. The evaluation of the affect of satellite SM data assimilation of AWRA-L estimation was not as clear as with the OzNet data, largely because of limited overlap period between cosmic-ray probe data availability and the model simulations (less than 24% overlap). Furthermore many of the CosmOz sites are in landscapes where the satellite soil moisture retrievals have higher error due to high vegetation cover fraction or proximity to coastal regions and water bodies and thus have less influence during the analysis state updating step. The work has demonstrated the suitability of the CosmOz network (and cosmic-ray probes in general) for verification of modelled moisture estimates of the root-zone over a wider range of biomes (from croplands to tropical rainforests) than is represented by the OzNet network, but also points to their potential to be assimilated into process models (e.g. Shuttleworth et al., 2013). Moreover, the probes can be situated in parts of the landscape where satellite moisture sensing techniques struggle to yield accurate SM retrievals (e.g. densely vegetated areas) and provide integrated profile SM estimation at a scale suited to land management practices.

5.6. Impact for water resources assessment and accounting A goal for data assimilation should be to improve model estimation when the model by itself is underperforming. This work has shown that the assimilation of satellite SM (albeit top-layer ‘‘observation’’) does improve soil water representation in AWRAL when the model performance is poor, particularly in the shallow root-zone region on the soil column. AWRA-L is coupled to river (AWRA-R) and groundwater (AWRA-G) model components of the larger AWRA system that is used for national scale water resource assessment and management. The interaction between the AWRAL data assimilation framework with the other model components, and thus the affect on the accuracy of wider water balance terms, remains to be explored. AWRA-L root-zone moisture analyses are useful as standalone products to evaluate trends in moisture patterns across Australia as part of a drought monitoring system (for example) or as independent verification of other land surface models. The next level of evaluation will be on the affect of the analysis soil water stores on other terms of the water balance model, particularly streamflow. While there is growing evidence of the improvements in soil water representation in models through satellite soil moisture assimilation, there are considerably fewer demonstrations of a flow-on positive impact on runoff estimation (e.g. Brocca et al., 2010). Improvement in soil water representation in models is itself a valuable outcome for water resources assessment and monitoring. Improvement or nil-impact in runoff estimation performance should be considered a good outcome. However degraded runoff estimation flow estimation should also be viewed as a positive outcome, as it points to potential inconsistency in the water balance model and lead to refinements, developments and improvements in the modelling system.

6. Summary and conclusion A continental soil moisture data assimilation framework was developed as part of the Australian Water Resources Assessment landscape (AWRA-L) modelling system. Perturbed meteorological forcing and the ensemble Kalman filter are used to assimilate AMSR-E and ASCAT satellite soil moisture (SM) retrievals into AWRA-L and generate cell-wise daily analyses of top-layer and

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

shallow root-zone soil moisture estimates (including the ensembles) at 0.05° resolution across Australia on a daily time step. AWRA-L model estimates of top-layer and shallow root-zone SM were evaluated against ground measurements, focusing on southeast Australia and the network of in situ probes across the Murrumbidgee catchment (OzNet), and across Australia through the emerging network of cosmic-ray proximal sensing technologies (CosmOz). AWRA-L model estimation of top-layer SM without the assimilation of satellite products (open-loop simulations) was generally quite good with average correlation of 0.75. This is likely due to the a priori calibration of the AWRA-L model against streamflow measurements, although the model is not calibrated against soil moisture data. The good performance of the AWRA-L model open-loop raised the bar for the assimilation to add value, and indeed at many of the OzNet sites we observed that the assimilation of satellite SM did not improve performance (sometimes degrading performance). However satellite data assimilation resulted in the most significant improvements in AWRA-L toplayer SM estimation for sites where the open-loop estimation was poorer than the satellite products. Satellite SM data assimilation into AWRA-L resulted in the most consistent improvements for the shallow root-zone 0–30 cm thickness. The assimilation of ASCAT SM improved AWRA-L estimation at most of the OzNet sites. And while the assimilation of AMSR-E degraded estimation at a small number of sites (4), the joint assimilation of both satellite products mitigated the degradation. The biggest adjustment (proportionally) to AWRA-L model states during the analysis update step occurred for the shallow-root soil layer. The analysis step adjusted the top-, shallow-root and deeproot soil layers, bring them into closer agreement with satellite SM. For the AWRA-L estimates of 0–90 cm SM, the inclusion of the deep-root layer in the open-loop root-zone calculation appeared to degrade performance at some OzNet sites. The success (or failure) of satellite SM for these sites was linked to the success (or failure) of assimilation on AWRA-L top-layer estimation. Overall, AWRA-L 0–90 cm root-zone estimation improved for 60% of the sites. We have demonstrated that the assimilation of AMSR-E and ASCAT SM products into the AWRA-L model can improve soil moisture estimation, particularly the shallow root-zone, when model performance (unconstrained) was poor. A desirable feature of any data assimilation system is that the observation and model errors are adequately specified such that the estimation is influenced by the optimal combination of model and observation, with the weight of influence shifting to the information source of highest reliability (lowest error). Continued effort is required in error characterisation of both the model and satellite products (including temporally varying observation erros), and maintaining high quality networks of ground sensing technology is critical to building confidence of the observation and model products particularly for water, and broader land, management and decision making. The CosmOz network, briefly explored in this work, has a clear and demonstrated role for model development and evaluation. Of key interest for regional studies is if these cosmic-ray-derived root-zone moisture measurements can be assimilated into process models (e.g. crop yield forecasting models) for improved land management practices. Acknowledgements This work was conducted under the Water Information Research and Development Alliance between CSIRO Water for a Healthy Country and the Bureau of Meteorology. We a grateful to our CSIRO colleagues Garth Warren, Tim Raupach and Dr. Edward King for sourcing and preparing the satellite data for this investigation. We also thank Aaron Hawdon, Dr. Lindsay Hutley,

2761

Steve Zegelin, Dr. Alessandra Monerris-Belda, Dr. Auro Almeida, and Dr. Matthew McCabe for allowing use of cosmic-ray probe data. The CosmOz network is funded by CSIRO Land and Water and CSIRO’s Water for a Healthy Country National Research Flagship. References Alsdorf, D.A., Lettenmaier, D.P., 2003. Tracking fresh water from space. Science 301, 1491–1494. Anderson, J., Anderson, S., 1999. A Monte Carlo implementation of the nonlinear filtering problem to produce ensemble assimilations and forecasts. Mon. Weather Rev. 127, 2741–2758. Anderson, M.C., Norman, J.M., Mecikalski, J.R., Otkin, J.A., Kustas, W.P., 2007. A climatological study of evapotranspiration and moisture stress across the continental United States based on thermal remote sensing: 1. Model formulation. J. Geophys. Res. 112 (D10117). http://dx.doi.org/10.1029/ 2006JD007506. Berry, P.A.M., Garlick, J.D., Freeman, E.L., Mathers, E.L., 2005. Global inland water monitoring from multi-mission altimetry. Geophys. Res. Lett. 32, L16401. http://dx.doi.org/10.1029/2005GL022814. Beven, K., Freer, J., 2001. Equifinality, data assimilation, and uncertainty in mechanistic modelling of complex environmental systems using the GLUE methodology. J. Hydrol. 249 (1–4), 11–29. Brocca, L., Melone, F., Moramarco, T., Wagner, W., Naeimi, V., Bartalis, Z., Hasenauer, S., 2010. Improving runoff prediction through the assimilation of the ASCAT soil moisture product. Hydrol. Earth Syst. Sci. 14 (10), 1881–1893. http://dx.doi.org/ 10.5194/hess-14-1881-2010. Caires, S., Sterl, A., 2003. Validation of ocean wind and wave data using triple collocation. J. Geophys. Res. 108 (C3), 3098. http://dx.doi.org/10.1029/ 2002JC001491. Chappell, A., Renzullo, L., Haylock, M., 2012. Spatial uncertainty to determine reliable daily precipitation maps. J. Geophys. Res. 117 (D17), D17115. http:// dx.doi.org/10.1029/2012JD017718. Chappell, A., Renzullo, L.J., Raupach, T.H., Haylock, M., 2013. Evaluating geostatistical methods of blending satellite and gauge data to estimate near real-time daily rainfall for Australia. J. Hydrol. 493, 105–114. http://dx.doi.org/ 10.1016/j.jhydrol.2013.04.024. Crow, W.T., Van Loon, E., 2006. Impact of incorrect model error assumptions on the sequential assimilation of remotely sensed surface soil moisture. J. Hydrol. 7 (3), 421–432. http://dx.doi.org/10.1175/JHM499.1. Crow, W.T., Van den Berg, M.J., 2010. An improved approach for estimating observation and model error parameters in soil moisture data assimilation. Water Resour. Res. 46 (12), 1–12. http://dx.doi.org/10.1029/2010WR009402. Crow, W.T., Wood, E.F., 2003. The assimilation of remotely sensed soil brightness temperature imagery into a land surface model using Ensemble Kalman filtering: a case study based on ESTAR measurements during SGP97. Adv. Water Res. 26. De Jeu, R.A.M., 2003. Retrieval of Land Surface Parameters Using Passive Microwave Observations, PhD thesis, Vrije Universiteit Amsterdam, Amsterdam, 120pp. Desilets, D., Zreda, M., 2013. Footprint diameter for a cosmic-ray soil moisture probe: theory and Monte Carlo simulations. Water Resour. Res. 49 (6), 3566– 3575. http://dx.doi.org/10.1002/wrcr.20187. Dharssi, I., Bovis, K., Macpherson, B., Jones, C., 2011. Operational assimilation of ASCAT surface soil wetness at the Met Office. Hydrol. Earth Syst. Sci. 8 (2), 4313–4354. http://dx.doi.org/10.5194/hessd-8-4313-2011. Donohue, R.J., McVicar, T.R., Roderick, M.L., 2009. Climate-related trends in Australian vegetation cover as inferred from stellite observations, 1981–2006. Glob. Change Biol. 15 (4), 1025–1039. Dorigo, W.A., Scipal, K., Parinussa, R.M., Liu, Y.Y., Wagner, W., de Jeu, R.A.M., Naeimi, V., 2010. Error characterisation of global active and passive microwave soil moisture datasets. Hydrol. Earth Syst. Sci. 14 (12), 2605–2616. http://dx.doi.org/ 10.5194/hess-14-2605-2010. Draper, C.S., Walker, J.P., Steinle, P.J., de Jeu, R.A.M., Holmes, T.R.H., 2009. An evaluation of AMSR-E derived soil moisture over Australia. Rem. Sens. Environ. 113 (4), 703–710. Draper, C., Mahfouf, J.-F., Calvet, J.-C., Martin, E., Wagner, W., 2011. Assimilation of ASCAT near-surface soil moisture into the SIM hydrological model over France. Hydrol. Earth Syst. Sci. 15 (12), 3829–3841. http://dx.doi.org/10.5194/hess-153829-2011. Draper, C.S., Reichle, R.H., De Lannoy, G.J.M., Liu, Q., 2012. Assimilation of passive and active microwave soil moisture retrievals. Geophys. Res. Lett. 39 (L04401), 2012. http://dx.doi.org/10.1029/2011GL050655. Evensen, G., 2003. The ensemble Kalman Filter: theoretical formulation and practical implementation. Ocean Dyn. 53 (4), 343–367. http://dx.doi.org/ 10.1007/s10236-003-0036-9. Franz, T.E., Zreda, M., Ferre, T.P.A., Rosolem, R., Zweck, C., Stillman, S., Zeng, X., Shuttleworth, W.J., 2012a. Measurement depth of the cosmic ray soil moisture probe affected by hydrogen from various sources. Water Resour. Res. 48 (8), W08515. http://dx.doi.org/10.1029/2012wr011871. Franz, T.E., Zreda, M., Rosolem, R., Ferre, T.P.A., 2012b. Field validation of a cosmicray neutron sensor using a distributed sensor network. Vadose Zone J. 11 (4). http://dx.doi.org/10.2136/vzj2012.0046.

2762

L.J. Renzullo et al. / Journal of Hydrology 519 (2014) 2747–2762

Gao, H., Tang, Q., Ferguson, C.R., Wood, E.F., Lettenmaier, D.P., 2010. Estimating the water budget of major US river basins via remote sensing. Int. J. Remote Sens. 31 (14), 3955–3978. http://dx.doi.org/10.1080/01431161.2010.483488. Guerschman, J.P., Van Dijk, A.I.J.M., Mattersdorf, G., Beringer, J., Hutley, L.B., Leuning, R., Pipunic, R.C., Sherman, B.S., 2009. Scaling of potential evapotranspiration with MODIS data reproduces flux observations and catchment water balance observations across Australia. J. Hydrol. 369 (1–2), 107–119. http://dx.doi.org/ 10.1016/j.jhydrol.2009.02.013. Hawdon, A., McJannet, D., Wallace, J., 2014. Calibration and correction procedures for cosmic-ray neutron soil moisture probes located across Australia. Water Resour. Res. 50 (6), 5029–5043. Henderson, B., Jin W., Renzullo, L., 2013. Error characterisation for dynamic data assimilation: soil moisture. CSIRO: Water for a Healthy Country National Research Flagship. Hou, A.Y., Skofronick-Jackson, G., Kummerow, C.D., Shepherd, J.M., 2008. Global precipitation measurement. In: Michaelides, S.C. (Ed.), Precipitation: Advances in Measurement, Estimation and Prediction. Springer-Verlag, pp. 131–170. Houtekamer, P., Mitchell, H., 1998. Data assimilation using an ensemble Kalman filter technique. Mon. Weather Rev. 126, 769–811. Huffman, G.J., Adler, R.F., Bolvin, D.T., Gu, G., Nelkin, E.J., Bowman, K.P., Hong, Y., Stocker, E.F., Wolff, D.B., 2007. The TRMM multisatellite precipitation analysis (TMPA): quasi-global, multiyear, combined-sensor precipitation estimates at fine scales. J. Hydromet. 8 (1), 38–55. http://dx.doi.org/10.1175/JHM560.1. Jones, D.A., Wang, W., Fawcett, R., Grant, I., 2007. Climate data for the Australian water availability project. In: Australian Water Availability Project Milestone Report. Bur. Met., Australia, 37pp. Jones, D.A., Wang, W., Fawcett, R., 2009. High-quality spatial climate data-sets for Australia, Australian Meteorol. Oceanogr. J. 58, 233–248. Joyce, R.J., Janowiak, J.E., Arkin, P.A., Xie, P., 2004. CMORPH: a method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydromet. 5, 487–503. Kalma, J.D., McVicar, T.R., McCabe, M.F., 2008. Estimating land surface evaporation: a review of methods using remotely sensed surface temperature data. Surveys Geophys. 29 (4–5), 421–469. http://dx.doi.org/10.1007/s10712-008-9037-z. King, E.A., Paget, M.J., Briggs, P.R., Trudinger, C.M., Raupach, M.R., 2009. Operational delivery of hydro-meteorological monitoring and modelling over the Australian continent. IEEE J. Sel. Top. Appl. Earth Observations Rem. Sens. 2 (4), 241–249. Koster, R.D., Guo, Z., Yang, R., Dirmeyer, P.A., Mitchell, K., Puma, M.J., 2009. On the nature of soil moisture in land surface models. J. Climate 22 (16), 4322–4335. http://dx.doi.org/10.1175/2009JCLI2832.1. Kumar, S.V., Reichle, R.H., Peters-Lidard, C.D., Koster, R.D., Zhan, X., Crow, W.T., Eylander, J.B., Houser, P.R., 2008. A land surface data assimilation framework using the land information system: description and applications. Adv. Water Res. 31 (11), 1419–1432. http://dx.doi.org/10.1016/j.advwatres.2008.01.013. Kumar, S.V., Reichle, R.H., Koster, R.D., Crow, W.T., Peters-Lidard, C.D., 2009. Role of subsurface physics in the assimilation of surface soil moisture observations. J. Hydromet. 10 (6), 1534–1547. http://dx.doi.org/10.1175/2009JHM1134.1. Liu, Y.Y., Parinussa, R.M., Dorigo, W.A., De Jeu, R.A.M., Wagner, W., van Dijk, A.I.J.M., Evans, J.P., 2011. Developing an improved soil moisture dataset by blending passive and active microwave satellite-based retrievals. Hydrol. Earth Syst. Sci. 15 (2), 425–436. http://dx.doi.org/10.5194/hess-15-425-2011. McCabe, M.F., Wood, E.F., Wojcik, R., Pan, M., Sheffield, J., Gao, H., Su, H., 2008. Hydrological consistency using multi-sensor remote sensing data for water and energy cycle studies. Rem. Sens. Environ. 112, 430–444. Miralles, D.G., Crow, W.T., Cosh, M.H., 2010. Estimating spatial sampling errors in coarse-scale soil moisture estimates derived from point-scale observations. http://dx.doi.org/10.1175/2010JHM1285.1. Mu, Q., Heinsch, F.A., Zhao, M., Running, S.W., 2007. Development of a global evapotranspiration algorithm based on MODIS and global meteorology data. Rem. Sens. Environ. 111 (4), 519–536. Njoku, E., Jackson, T., Lakshmi, V., Chan, T., Nghiem, S., 2003. Soil moisture retrieval from AMSR-E. IEEE Trans. Geosci. Rem. Sens. 41, 215–229. Owe, M., de Jeu, R., Holmes, T., 2008. Multisensor historical climatology of satellitederived global land surface moisture. J. Geophys. Res. 113 (F01002), 1–17. http://dx.doi.org/10.1029/2007JF000769. Pan, M., Sahoo, A.K., Troy, T.J., Vinukollu, R.K., Sheffield, J., Wood, E.F., 2012. Multisource estimation of long-term terrestrial water budget for major global river basins. J. Climate 25 (9), 3191–3206. http://dx.doi.org/10.1175/JCLI-D-1100300.1. Parrens, M., Mahfouf, J.-F., Barbu, A.L., Calvet, J.-C., 2014. Assimilation of surface soil moisture into a multilayer soil model: design and evaluation at local scale. Hydrol. Earth Syst. Sci. 18 (2), 673–689. http://dx.doi.org/10.5194/hess-18-6732014. Perraud, J.-M., Collins, D., Bowden, J.C., Raupach, T., Manser, P.A., Stenson, M.P., Renzullo, L.J., 2013. A balancing act in heterogeneous computing – developing the AWRA-Landscape data assimilation system. In: Proc. MODSIM2013. Pipunic, R.C., Walker, J.P., Western, A., 2008. Assimilation of remotely sensed data for improved latent and sensible heat flux prediction: a comparative synthetic study. Rem. Sens. Environ. 112 (4), 1295–1305. http://dx.doi.org/10.1016/ j.rse.2007.02.038. Pipunic, R.C., Walker, J.P., Western, A.W., Trudinger, C.M., 2013. Assimilation of multiple data types for improved heat flux prediction: a one-dimensional field study. Rem. Sens. Environ. 136, 315–329. http://dx.doi.org/10.1016/ j.rse.2013.05.015.

Raupach, M.R., Trudinger, C.M., Briggs, P.R., Hverd, V., King, E.A., Paget, M.J., 2009. 108 years of Australian water balance. Hydrol. Earth Syst. Sci. Reichle, R.H., Koster, R.D., 2004. Bias reduction in short records of satellite soil moisture. Geophys. Res. Lett. 31 (19), 2–5. http://dx.doi.org/10.1029/ 2004GL020938. Reichle, R., Koster, R., 2005. Global assimilation of satellite surface soil moisture retrievals into the NASA Catchment Land Model. Geophys. Res. Lett. 32, L02404. http://dx.doi.org/10.1029/2004GL021700. Reichle, R.H., McLaughlin, D.B., Entekhabi, D., 2002. Hydrologic data assimilation with the ensemble Kalman filter. Mon. Weather Rev. 130 (1), 103–114. Reichle, R.H., Koster, R.D., Liu, P., Mahanama, S.P.P., Njoku, E.G., Owe, M., 2007. Comparison and assimilation of global soil moisture retrievals from the advanced microwave scanning radiometer for the earth observing system (AMSR-E) and the scanning multichannel microwave radiometer (SMMR). J. Geophys. Res. 112 (D9), D09108. http://dx.doi.org/10.1029/2006JD008033. Reichle, R.H., Crow, W.T., Keppenne, C.L., 2008. An adaptive ensemble Kalman filter for soil moisture data assimilation. Water Resour. Res. 44 (3), 1–13. http:// dx.doi.org/10.1029/2007WR006357. Renzullo, L., Chappell, A., Raupah, T., Dyce, P., Li, M., Shao, Q., 2011. An assessment of blended satellite-gauge precipitation products for Australia, CSIRO: Water for a Healthy Country National Research Flagship . Rodell, M., Chen, J., Kato, H., Famiglietti, J.S., Nigro, J., Wilson, C.R., 2006. Estimating groundwater storage changes in the Mississippi River basin (USA) using GRACE. Hydrogeol. J. 15 (1), 159–166. http://dx.doi.org/10.1007/s10040-006-0103-7. Rodell, M., Velicogna, I., Famiglietti, J.S., 2009. Satellite-based estimates of groundwater depletion in India. Nature 460 (7258), 999–1002. http:// dx.doi.org/10.1038/nature08238. Scipal, K., Holmes, T., de Jeu, R., Naeimi, V., Wagner, W., 2008. A possible solution for the problem of estimating the error structure of global soil moisture data sets. Geophys. Res. Lett. 35, L24403. http://dx.doi.org/10.1029/2008GL035599. Sheffield, J., Ferguson, C.R., Troy, T.J., Wood, E.F., McCabe, M.F., 2009. Closing the terrestrial water budget from satellite remote sensing. Geophys. Res. Lett. 36 (7), 1–5. http://dx.doi.org/10.1029/2009GL037338. Shuttleworth, J., Rosolem, R., Zreda, M., Franz, T., 2013. The cosmic-ray soil moisture interaction code (COSMIC) for use in data assimilation. Hydrol. Earth Syst. Sci. 10, 1097–1125. Smith, A.B., Walker, J.P., Western, A.W., Young, R.I., Ellett, K.M., Pipunic, R.C., Grayson, R.B., Siriwardena, Chiew, F.H.S., Richter, H., 2012. The Murrumbidgee soil moisture monitoring network data set. Water Resour. Res. 48 (7), 1–6. http://dx.doi.org/10.1029/2012WR011976. Stenson, M.P., Fitch, P., Vleeshouwer, J., Frost A.J., Bai, Q., Lerat, J., Leighton, B., Knapp, S., Warren, G., Van Dijk, A.I.J.M., Bacon, D., Pena-Arancibia, J. L., Manser, P., Shoesmith, J., 2011. Operationalising the Australian water resources assessment system. In: Proceedings Water Information Research and Development Alliance Science Symposium, Melbourne 1–5 August 2011, pp. 36–45. Su, C.-H., Ryu, D., Young, R.I., Western, A.W., Wagner, W., 2013. Inter-comparison of microwave satellite soil moisture retrievals over the Murrumbidgee Basin, southeast Australia. Rem. Sens. Environ. 134 (2013), 1–11. http://dx.doi.org/ 10.1016/j.rse.2013.02.016. Turner, M., Walker, J., Oke, P., 2008. Ensemble member generation for sequential data assimilation. Rem. Sens. Environ. 112 (4), 1421–1433. http://dx.doi.org/ 10.1016/j.rse.2007.02.042. Van Dijk, A.I.J.M., 2010. The Australian Water Resources Assessment System; Technical Report 3. Landscape Model (version 0.5) Technical Description. WIRADA/Water for a Healthy Country Flagship, Canberra. Van Dijk, A.I.J.M., Marvanek, S., 2010. Derivation of a simple soil drainage model. AWRA Background paper 2010/1. WIRADA/Water for a Healthy Country Flagship, Canberra. Van Dijk, A.I.J.M., Renzullo, L.J., 2011. Water resource monitoring systems and the role of satellite observations. Hydrol. Earth Syst. Sci. 15 (1), 39–55. http:// dx.doi.org/10.5194/hess-15-39-2011. Van Dijk, A.I.J.M., Renzullo, L.J., Rodell, M., 2011. Use of gravity recovery and climate experiment terrestrial water storage retrievals to evaluate model estimates by the Australian water resources assessment system. Water Resour. Res. 47 (11), 1–12. http://dx.doi.org/10.1029/2011WR010714. Viney, N. R., Van Dijk, A.I.J.M., Vaze, J., 2011. Comparison of models for estimating spatial patterns of streamflow across Australia. In: Proceedings Water Information Research and Development Alliance Science Symposium, Melbourne 1–5 August 2011, pp. 30–35. Wagner, W., Lemoine, G., Rott, H., 1999. A method for estimating soil moisture from ERS scatterometer and soil data. Rem. Sens. Environ. 70, 191–207. Wood, E.F., Roundy, J.K., Troy, T.J., van Beek, L.P.H., Bierkens, M.F.P., Blyth, E., de Roo, A., Ek, M., Famiglietti, J., Gochis, D., van de Giesen, N., Houser, P., Jaffe, P.R., Kollet, S., Lehner, B., Lettenmaier, D.P., Peters-Lidard, C., Sivapalan, M., Sheffield, J., Wade, A., Whitehead, P., 2011. Hyperresolution global land surface modeling: meeting a grand challenge for monitoring Earth’s terrestrial water. Water Resour. Res. 47 (5), W05301. http://dx.doi.org/10.1029/2010WR010090. Zreda, M., Shuttleworth, W.J., Zeng, X., Zweck, C., Desilets, D., Franz, T., Rosolem, R., 2012. COSMOS: the cosmic-ray soil moisture observing system. Hydrol. Earth Syst. Sci. 16, 4079–4099. http://dx.doi.org/10.5194/hess-16-1-2012. Zwieback, S., Scipal, K., Dorigo, W., Wagner, W., 2012. Structural and statistical properties of the collocation technique for error characterization. Nonlin. Process. Geophys. 19, 69–80. http://dx.doi.org/10.5194/npg-19-69-2012.