A land surface data assimilation framework using the land information system: Description and applications.

Abstract

The Land Information System (LIS) is an established land surface modeling framework that integrates various community land surface models, ground measurements, satellite-based observations, high performance computing and data management tools. The use of advanced software engineering principles in LIS allows interoperability of individual system components and thus enables assessment and prediction of hydrologic conditions at various spatial and temporal scales. In this work, we describe a sequential data assimilation extension of LIS that incorporates multiple observational sources, land surface models and assimilation algorithms. These capabilities are demonstrated here in a suite of experiments that use the ensemble Kalman filter (EnKF) and assimilation through direct insertion. In a soil moisture experiment, we discuss the impact of differences in modeling approaches on assimilation performance. Provided careful choice of model error parameters, we find that two entirely different hydrological modeling approaches offer comparable assimilation results. In a snow assimilation experiment, we investigate the relative merits of assimilating different types of observations (snow cover area and snow water equivalent). The experiments show that data assimilation enhancements in LIS are uniquely suited to compare the assimilation of various data types into different land surface models within a single framework. The high performance infrastructure provides adequate support for efficient data assimilation integrations of high computational granularity.

Full text

Page 1

A land surface data assimilation framework using the land
information system: Description and applications
Sujay V. Kumara,b,*, Rolf H. Reichlea,c, Christa D. Peters-Lidardb, Randal D. Kosterc,
Xiwu Zhand, Wade T. Crowe, John B. Eylanderf, Paul R. Houserg
aUniversity of Maryland at Baltimore County, Goddard Earth Sciences and Technology Center, Baltimore, MD 21250, United States
bHydrological Sciences Branch, NASA Goddard Space Flight Center, Greenbelt, MD 20771, United States
cNASA Global Modeling and Assimilation Office, NASA Goddard Space Flight Center, Greenbelt, MD 20771, United States
dNOAA-NESDIS Center for Satellite Applications and Research, Camp Springs, MD 20746, United States
eUSDA-ARS Hydrology and Remote Sensing Laboratory, Beltsville, MD 20705, United States
fAir and Space Models Integration Branch, HQ Air Force Weather Agency, AFB, NE 68113, United States
gCenter for Research on Environment and Water, George Mason University, Calverton, MD 20705, United States
Received 17 May 2007; received in revised form 19 December 2007; accepted 20 January 2008
Available online 6 February 2008
Abstract
The Land Information System (LIS) is an established land surface modeling framework that integrates various community land sur-
face models, ground measurements, satellite-based observations, high performance computing and data management tools. The use of
advanced software engineering principles in LIS allows interoperability of individual system components and thus enables assessment
and prediction of hydrologic conditions at various spatial and temporal scales. In this work, we describe a sequential data assimilation
extension of LIS that incorporates multiple observational sources, land surface models and assimilation algorithms. These capabilities
are demonstrated here in a suite of experiments that use the ensemble Kalman filter (EnKF) and assimilation through direct insertion. In
a soil moisture experiment, we discuss the impact of differences in modeling approaches on assimilation performance. Provided careful
choice of model error parameters, we find that two entirely different hydrological modeling approaches offer comparable assimilation
results. In a snow assimilation experiment, we investigate the relative merits of assimilating different types of observations (snow cover
area and snow water equivalent). The experiments show that data assimilation enhancements in LIS are uniquely suited to compare the
assimilation of various data types into different land surface models within a single framework. The high performance infrastructure pro-
vides adequate support for efficient data assimilation integrations of high computational granularity.
? 2008 Elsevier Ltd. All rights reserved.
Keywords: Land surface modeling; Data assimilation; Remote sensing; Hydrology; Soil moisture; Snow
1. Introduction
The land surface has a profound influence on regional
and global weather and climate through the exchanges of
moisture and energy between the soil, vegetation and snow-
pack with the overlying atmosphere. Therefore, a realistic
characterization of the land surface moisture and energy
stores that control these exchanges is important for
improving our understanding and prediction of land–
atmospheric interactions. An accurate characterization of
the land surface can lead to improvements not only in
weather and climate prediction, but also in other applica-
tions such as hazard mitigation (floods and droughts), agri-
cultural production and water resources management. The
accuracy of predictions from a Land Surface Model (LSM)
depends on the model’s representation of physical pro-
cesses, the quality of the model inputs and forcings and
the accuracy of the model parameters. Constraining the
model predictions with observations via data assimilation
0309-1708/$ - see front matter ? 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advwatres.2008.01.013
*Corresponding author. Tel.: +1 301 286 8663; fax: +1 301 286 8624.
E-mail address: Sujay.V.Kumar@nasa.gov (S.V. Kumar).
www.elsevier.com/locate/advwatres
Available online at www.sciencedirect.com
Advances in Water Resources 31 (2008) 1419–1432

Page 2

methods is an effective way to attenuate model errors and
improve the model’s predictive skills.
Data assimilation has been used in many scientific appli-
cations [25] as a way to improve deterministic model accu-
racy. Historically, the use of data assimilation techniques in
the Earth sciences has been restricted primarily to meteoro-
logical and oceanographic applications. In recent years,
many researchers have developed data assimilation tech-
niques to exploit the increased availability of remotely
sensed land surface variables [26,29,43]. A rapidly growing
number of studies evaluate the assimilation of soil mois-
ture, snow and surface skin temperature observations
(e.g., [1,3,5,11,24,33,35,38,39,41,42]). These studies not
only demonstrate the potential of data assimilation to
improve land surface predictions, but also describe the dif-
ficulties in managing the complexities associated with data
handling, computational burdens and the associated trade-
offs of various algorithms. Such difficulties are especially
apparent when large amounts of remotely sensed observa-
tions of land surface states are assimilated.
As noted by [43], a comprehensive software framework
is required that integrates physical models, observations,
assimilation algorithms, and the necessary computing
infrastructure to address the complexities associated with
land surface data assimilation. The present article describes
the development of such a framework as an extension of
the existing Land Information System (LIS). As described
by [22], LIS is a land modeling system that operates several
community land surface models with the required initial
and boundary conditions. LIS is designed with advanced
software engineering principles and it provides many user
extensible interfaces to incorporate diverse data sets from
different sources as inputs to the LSMs. LIS also includes
generic support for high performance computing, enabling
the use of LSMs at global scales with spatial resolutions as
high as 1 km. LIS has also been coupled to regional numer-
ical weather prediction models to further enable the inves-
tigation of land–atmosphere interactions [21].
The LIS data assimilation extension described here is
designed to be a generic, sequential component. Sequential
assimilation algorithms step recursively through time,
alternating between a model propagation step and a data
assimilation update step. Examples of such algorithms
are Direct Insertion (DI), the Extended Kalman filter
(EKF), particle filters and the Ensemble Kalman filter
(EnKF), each of which has been applied successfully to
hydrologic data assimilation [36,38,44,45]. A sequential
structure is also convenient for processing measurements
in real time, such as in an operational setting. To date,
the research nature of land surface data assimilation has
limited most studies to the assimilation of a single measure-
ment type into a single model using a single methodology.
The LIS assimilation extension is designed to facilitate the
much needed cross-comparison studies that evaluate the
interoperable and integrated use of several kinds of assim-
ilation algorithms, observations data sets and land surface
models.
The paper is organized as follows. After a brief review of
the Kalman filter (Section 2), we describe the design of the
LIS interoperable data assimilation component (Section 3).
The capabilities enabled by the design are illustrated with
two sets of synthetic experiments that provide new insights
in their own right. The first set of experiments investigates
the use of different land surface models for soil moisture
estimation (Section 4.1). In the second set, we compare
the assimilation of two different types of snow observations
(Section 4.2). Diagnostics of the assimilation performance
and computational aspects are discussed in Section 5. Con-
clusions and a summary of future planned assimilation
extensions to LIS are given in Section 6.
2. Sequential data assimilation
Sequential assimilation algorithms step recursively
through time, alternating between a model propagation
step and a data assimilation update step. The latter occurs
whenever observations are available. In such a progression,
the most recent updates reflect the accumulated informa-
tion from all the observations up to this time. Here, we
represent the nonlinear land surface model in the generic
form
xkþ1¼ fkðxk;wkÞ;
where xk represents the state vector at time k, fð?Þ is the
nonlinear model forward operator, and wk represents the
uncertainties due to errors in the model formulation and
uncertain model forcings. The observations at time k þ 1
denoted by ykþ1are connected to the system states via the
observation operator hkþ1:
ykþ1¼ hkþ1ðxkþ1;vkþ1Þ;
where measurement errors are represented by vk. The mod-
el and observation errors wkand vkare typically assumed
to be independent Gaussian random vectors with mean
zero and covariances Qkand Rk, respectively, that are also
uncorrelated in time.
The propagation step of the sequential assimilation
algorithm consists of integrating equation (1) (from an ini-
tial state estimate ^ xþ
model forecast ^ x?
new state estimate ^ xþ
ð1Þ
ð2Þ
kat time k and with wk¼ 0) to give a
kþ1at time k þ 1. In the update step, a
kþ1(also known as the analysis),
kþ1þ Kðykþ1? hkþ1ð^ x?
is computed based on the gain matrix K and on the inno-
vations vector (ykþ1? hkþ1ð^ x?
tween the measurements and the model’s prediction of
the observations).
The gain matrix is generally a function of the model and
observation error covariances. Its computation is the pri-
mary difference between the different sequential assimila-
tion algorithms.
In the Kalman filter family of algorithms the error char-
acteristics of the model are dynamically evolved in time,
and for linear systems with error characteristics as specified
^ xþ
kþ1¼ ^ x?
kþ1ÞÞ;
ð3Þ
kþ1Þ; that is, the difference be-
1420
S.V. Kumar et al./Advances in Water Resources 31 (2008) 1419–1432

Page 3

above, the gain is optimal (in the sense of minimum estima-
tion error variance).
Though the LIS data assimilation framework accommo-
dates different flavors of Kalman filters, we focus primarily
on the use of the EnKF in this article. The salient feature of
the EnKF is the approximation of model and forcing error
covariances through the propagation of an ensemble of
model trajectories. Each ensemble member is subject to a
different realization of model and forcing errors. Each
ensemble member could also employ different sets of model
parameters or even entirely different land surface models,
although such an approach is not used in this paper. After
the assimilation update, the model is again evolved forward
from the analysis state to the next observation time and the
process is repeated [36].
3. Design of the LIS data assimilation component
The LIS software architecture employs object oriented
programming paradigms to enable a modular and extensi-
ble system [22]. All functional pieces in LIS are imple-
mented asextensiblecomponents,
meteorological input schemes, sources of land surface
parameters, modeling domains and running modes, follow-
ing the principles of object oriented frameworks [2,28]. This
plug-and-play design has enabled the inclusion of several
user-defined extensions of each of these functional abstrac-
tions. The addition of data assimilation capabilities follows
a similar extensible design.
The data assimilation extension in LIS is designed to be
a sequential operator, where simulation variables are cor-
rected at every observation time. Three main abstractions
encapsulate the overall behavior of the data assimilation
process defined in LIS (Fig. 1). These abstractions repre-
sent: (1) the data assimilation algorithm, (2) the observa-
tions, and (3) the land surface models. The data
assimilation algorithm abstraction represents the specific
algorithm that is used for the assimilation update, for
example direct insertion or the EnKF. Various sources of
observational data and corresponding measurement mod-
includingLSMs,
els that provide data to be assimilated are represented by
the Observations abstraction. Examples of hydrological
remote sensing data include surface soil moisture and snow
water equivalent (SWE) retrievals from the Advanced
Microwave Scanning Radiometer for the Earth Observing
System (AMSR-E) as well as snow cover observations from
the Moderate Resolution Imaging Spectroradiometer
(MODIS). Finally, the Land Surface Models abstraction
represents various LSM implementations in LIS (such as
the Noah LSM and the NASA Catchment LSM described
below), along with the associated model parameters and
surface meteorological forcing data.
The LIS core, which is the primary software that enables
the integrated use of various extensible components in LIS,
combines the use of these data assimilation abstractions.
The abstractions defined by the LIS core need to be
extended for each specific data assimilation instance. Data
are exchanged between the various LIS components
through the constructs of the Earth System Modeling
Framework (ESMF, [18]. ESMF provides a standardized,
self-describing format for data exchange between the sys-
tem components shown in Fig. 1 through objects of the
ESMF State datatype.
Additional illustration of the sequential assimilation is
provided in Fig. 2, which shows the interaction of the data
assimilation abstractions for a single cycle of a typical
sequential data assimilation operation (using the notation
of Section 2). The vertical axis represents the sequence of
operations and the horizontal axis represents the interact-
ing modules in LIS as described above (Fig. 1). During a
typical cycle (from time k to k þ 1), the forecast step is per-
formed first to project the previous state (and possibly the
error covariances) forward in time (from ^ xþ
obtain the a priori estimates for the current timestep. Next,
the observations to be assimilated by the Observation
Module are read and packaged as an ESMF State object,
represented by the OBS State ykþ1. This is followed by the
measurement update (or analysis) step that incorporates
the new measurements into the a priori estimate ^ x?
obtain an improved a posteriori estimate ^ xþ
by the updated LSM State. This cycle repeats until the end
of the simulation. Additional components and interactions
are added to simulate the uncertainty and error propaga-
tion in specific instances of a sequential data assimilation
algorithm such as the EnKF. For example, a Perturba-
tion Module is included for the EnKF to control the evo-
lution of model states and error characteristics. For this
study, we implemented the EnKF and perturbation soft-
ware developed by [33] as LIS modules.
The ESMF data objects that enable the exchange of
information between data assimilation components contain
not only the data but also certain self-describing metadata.
For example, the LSM State object contains the prognostic
state variables to be assimilated. It also contains attributes
describing the data such as the names, units, sign conven-
tions and maximum and minimum acceptable values. Sim-
ilarly, the OBS State object contains the observation data
kto ^ x?
kþ1) and
kþ1to
kþ1, represented
LIS CORE
Land Surface
Models
Noah
CLM
Catchment
....
Data Assimilation
Algorithms
DI
EnKF
....
Observations
Soil Moisture
Snow
Temperature
........
Fig. 1. Data assimilation abstractions in LIS.
S.V. Kumar et al./Advances in Water Resources 31 (2008) 1419–1432
1421

Page 4

and several attributes that describe them. These objects are
queried to retrieve the data and the metadata by data
assimilation specific segments of the framework. In other
words, the observations and the relevant LSM prognostic
variablesmustbepackaged
ESMF State objects to select a specific data assimilation
instance in LIS. The use of ESMF State objects thus pro-
vides a flexible representation to describe the data being
exchanged.
asselfdescribing
4. Data assimilation experiments and results
The implementation of data assimilation extensions in
LIS facilitates an advanced system that enables the interop-
erable use of multiple data assimilation algorithms, land
surface models and observations in a single framework.
In addition, the existing support for computational paral-
lelism enables the use of the system for computationally
demanding applications. This section presents two sets of
data assimilation experiments that demonstrate these capa-
bilities. The first set of experiments investigates the use of
different land surface models for soil moisture estimation
(Section 4.1). In the second set, we compare the assimila-
tion of two different types of snow observations (Section
4.2). Each type of snow observation is assimilated using
an appropriate assimilation algorithm.
To demonstrate the interoperability of land surface
models in a data assimilation mode, two different LSMs
are used in the experiments: the Noah LSM [14] and the
NASA Catchment LSM [20]. Both models dynamically
predict land surface water and energy fluxes, but the model
structures and parameterizations differ. For example, the
Catchment LSM employs a topographically based hydro-
logic catchment representation with an explicit treatment
of subgrid soil moisture variability and its effect on runoff
and evaporation. By contrast, Noah uses a layer-based
approach to soil moisture modeling. The Catchment
LSM and the Noah LSM include physically based but dis-
tinct snow sub-models to represent processes related to
snowpack evolution. The Catchment LSM uses three snow
layers, whereas Noah uses only one snow layer. The two
types of snow observations used in the second set of exper-
iments are snow-cover area (SCA) and SWE.
The two sets of synthetic experiments are designed to
demonstrate the sensitivity of the assimilation process to
model parameterizations and physical representations.
They also provide an opportunity to compare and evaluate
the relative merits of assimilating different data types with
different assimilation algorithms within a single frame-
work. The basic structure of the experiments is as follows:
First, the model is integrated to obtain the assumed ‘‘true”
state of the land surface, referred to as the control (or
truth) run. The observations to be assimilated are then gen-
erated from the control run’s outputs by introducing real-
isticretrievalerrors.Degraded
simulations are conducted using uncertain inputs to
degrade the model estimates. The uncertainty in the inputs
is introduced by using a meteorological forcing data set for
the time period that is different from that of the truth run
(e.g. as produced by a different data provider), primarily to
ensure that the errors in the open loop simulations are real-
istic. Finally, the assimilation integrations are conducted
by assimilating the synthetic observations into the open
loop simulation for each grid cell independently (also
or‘‘openloop”
LIS
Simulation
k
Data Assimilation
Module
Observations
Module
LSM
Module
OBS_State ()
LSM_State ( )
LSM_State ()
k+1
Run land surface
model (Forecast)
Read Observations
Analysis
Fig. 2. Sequence of component interactions for a sequential data assimilation cycle.
1422
S.V. Kumar et al./Advances in Water Resources 31 (2008) 1419–1432

Page 5

known as ‘‘one-dimensional” assimilation, [34]). Since the
true fields are known in these experiments, we can easily
assess the performance of the assimilation approaches.
4.1. Soil moisture experiment
Soil moisture has a profound influence on the temporal
and spatial variability of weather and climate conditions
because it plays an important role in controlling the
exchange of water and energy between the land surface
and the atmosphere through evaporation and plant tran-
spiration. Surface soil moisture observations can be
obtained from measurements of microwave radiation emit-
ted by the land surface. Hydrologic data assimilation
efforts have focused on improving soil moisture estimates
by incorporating such remote sensing data. In this section,
we describe a suite of experiments that simulate the assim-
ilation of remotely-sensed surface soil moisture into the
Catchment and Noah LSMs.
Moving from the surface downward, the vertical subsur-
face representation in the Noah LSM includes four soil lay-
ers: a 10-cm thick surface layer, a 30-cm thick root zone
layer, a 60-cm thick deep root zone layer, and a 1-m thick
sub-root zone layer. The model keeps track of eight prog-
nostic soil moisture variables, two for each layer (the total
volumetric soil moisture and the liquid fraction of the vol-
umetric soil moisture content). Noah employs the vertically
integrated Richards equation [37] to compute the transfer
of water within the soil profile.
In contrast, the Catchment LSM uses three bulk mois-
ture variables and topographic characteristics to diagnose
subsurface moisture for each catchment (or computational
unit). These non-traditional bulk moisture variables repre-
sent equilibrium conditions associated with the water table
distribution and the non-equilibrium conditions in the root
zone and near-surface. The Catchment LSM explicitly
computes the sub-catchment spatial distribution of mois-
ture based on topography, subdividing the catchment into
three distinct moisture regimes [12]. Soil moisture in a 2-cm
surface layer and in a 1-m root zone layer (which includes
the surface layer) are diagnosed from the bulk moisture
variables.
In both land surface models, soil moisture responds sim-
ilarly to the meteorological forcing, with soil layers closer
to the surface having a faster response and the fluctuations
in the deeper layers becoming less variable. The character-
istic response time scales and the vertical soil moisture cou-
pling of soil moisture are different for the two models. In
the following, surface soil moisture for Catchment refers
to its 2-cm surface layer and surface soil moisture for Noah
refers to its top 10-cm layer. For the Catchment LSM, root
zone soil moisture refers to the model’s 1-m root zone
layer. For Noah, an equivalent 1-m thick layer is computed
as the weighted average of Noah’s top three layers and is
henceforth referred to as root zone soil moisture.
The modeling domain for the soil moisture experiment
covers roughly the Continental United States (CONUS,
from 30.5?N,124.5?W to 50.5?N,75.5?W) at 1? spatial reso-
lution. The two land models are spun up from January 1,
2000 using surface meteorological forcings from the Global
Data Assimilation System (GDAS; the global operational
weather forecast model of the National Center for Envi-
ronmental Prediction [9]). The eight months from April 1
to December 1, 2003 are used as the experiment period.
‘‘True” land surface conditions are simulated by integrat-
ing each model with GDAS forcing. Synthetic observations
of surface soil moisture for each model are generated from
the respective truth integrations by simulating retrieval
errors typically associated with soil moisture products from
microwave sensors. To account for difficulties in retrieving
soil moisture under dense vegetation canopies, the syn-
thetic surface soil moistures are masked out for both exper-
iments when Green Vegetation Fraction values (used in
Noah) exceed 0.7. The limitations of soil moisture retri-
evals in the presence of precipitation or snow are also sim-
ulated by introducing data masking for these events.
Moreover, random Gaussian noise with an error standard
deviation of 0.03 m3m?3(volumetric soil moisture) is
added to the synthetic observations to mimic measurement
uncertainties. The magnitude of these perturbations is an
optimistic estimate of error levels in future surface soil
moisture retrievals from space-borne L-band radiometers.
Next, open loop simulations are conducted by forcing the
land surface models with surface meteorological forcing
from the Goddard Earth Observing System (GEOS; [27]).
In other words, model errors are represented by the differ-
ence between the GDAS and the GEOS forcing data sets.
The synthetic observations from the Catchment LSM
are then assimilated into the Catchment LSM open loop
model (using GEOS forcing) with the EnKF, once a day
at 12Z. Likewise, Noah retrievals are assimilated into the
Noah model. Perturbations on meteorological forcing
inputs, model prognostic variables and observations are
applied to maintain an ensemble of land surface conditions
that represent the uncertainty in soil moisture states [33],
with details shown in Table 1. Zero-mean, normally dis-
tributed additive perturbations were applied to the down-
ward longwave radiation and near surface temperature
forcing, and log-normally distributed multiplicative pertur-
bations (with mean 1) are applied to precipitation and
downward shortwave radiation forcing. Furthermore, the
Catchment LSM prognostic variables catchment deficit
and surface excess are perturbed with additive noise. For
the Catchment LSM, the surface soil moisture observations
are used to adjust all three bulk soil moisture prognostic
variables (catchment deficit, root zone excess and surface
excess) in the EnKF update step. Similarly, the Noah prog-
nostic variables related to all four soil layers are updated in
response to the surface soil moisture observations. Vertical
correlations are imposed on the perturbations for the Noah
soil moisture prognostic variables for individual layers.
To assess the performance of the assimilation integra-
tions, root mean square errors (RMSEs) with respect to
the true simulation are computed for the surface and root
S.V. Kumar et al./Advances in Water Resources 31 (2008) 1419–1432
1423