Plain Language Summary
Soil moisture is an essential variable coupling the land surface to the
atmosphere. Accurate estimates of soil moisture are important for
predicting where clouds will form, assessing drought and fire weather
risks, and assisting with decisions for agricultural production. There
are multiple estimates of soil moisture available, and in this study, we
compare three different types of soil moisture estimates from the
National Oceanic and Atmospheric Administration (NOAA), including two
types of soil moisture models and direct observations, as well as direct
observations from the United States Department of Agriculture. Both
models have soil moisture values that are too wet in dry regions of the
United States and too dry in wet regions. The modeled soil moisture also
does not change as rapidly from day-to-day as they do in the direct
observations of soil moisture. Further analysis at different soil depths
shows which depths have the largest differences between these three soil
moisture estimates. Understanding these NOAA soil moisture estimate
differences is important for forecasters and decision makers and can be
useful for the further development of atmospheric and land surface
models.
1 Introduction
Knowledge of soil moisture is essential for many Earth system
applications, such as forecasting cloud formation (e.g., Ek and
Holtslag, 2004), monitoring drought, flood and fire risks (e.g., Svoboda
et al., 2002; Rigden et al., 2020), and providing instrumental
information for agricultural production (e.g., Madadgar et al., 2017).
As such, several advancements in the estimation and utilization of soil
moisture have recently transpired. For example, in efforts to improve
the accuracy of numerical weather prediction (NWP) and climate models,
model developers have focused on increasing the coupling between the
land surface and atmosphere components of their data assimilation
systems to eliminate persistent atmospheric prediction biases (e.g.,
Benjamin et al., 2022). Furthermore, NWP models are also beginning to
explore the direct assimilation of new soil moisture observations (e.g.,
Carrera et al., 2019; Muñoz-Sabater et al., 2019; Lin and Pu, 2020).
For these reasons, it is critical to understand the differences in the
available soil moisture estimates from models or observations that are
used in science, forecasting or agricultural applications. Different
estimates of soil moisture exist across the continental United States
(CONUS), each with its own benefits and shortfalls. For example, in situ
observations are often considered to be the most accurate and are
therefore used as a benchmark. However, they have limited spatial
coverage (e.g., Quiring et al., 2016). Other products, such as those
from low Earth orbiting satellites, have lower temporal and spatial
resolution (e.g., Liu et al., 2016). Soil moisture estimates from models
depend on many assumptions and reflect the influence of
observation-based data to different degrees (e.g., Smirnova et al.,
1997; Huang et al., 1996; Mitchell et al., 2004). Several recent studies
have made significant progress in comparing many of the available soil
moisture estimates. Some studies have compared soil moisture temporal
variability and memory between many large-scale land surface models
(LSMs) and in situ soil monitoring networks across the CONUS, noting
certain biases and uncertainties in the various estimates (e.g., Robock
et al., 2003; Xia et al., 2014; 2015a; Dirmeyer et al., 2016). Recent
studies have extended soil moisture comparisons to include soil moisture
retrievals from new satellite platforms (e.g., Shellito et al., 2016;
Pan et al., 2016; Ford and Quiring, 2019).
While various LSMs have been compared to in situ observations for
several decades, these prior studies have primarily focused on models
with horizontal resolutions on the scales of ⅛ degree or larger (i.e.,
the North American Land Data Assimilation System models; Mitchell et
al., 2004; Xia et al., 2012). Given the local, mesoscale variability of
soil moisture process and the subsequent impacts of soil moisture on
atmospheric prediction (e.g., Koster et al., 2004; Taylor et al., 2011),
high-resolution models should also be tested. Recently, Min et al.
(2021) compared near-surface atmospheric and soil variables from the
High-Resolution Rapid Refresh (HRRR) model (Dowell et al., 2022; James
et al., 2022), which has horizontal grid spacing of 3 km, to
observations from the New York State Mesonet. They found that soil
moisture was underestimated in HRRR, which contributed to warm and dry
biases in atmospheric forecasts. However, how does soil moisture from
the HRRR model (NOAA’s current operational, convection-allowing model)
compare to soil moisture estimates across all of CONUS?
In this study, we focus on a comparison of soil moisture estimates from
NOAA. In particular, we uniquely include soil moisture estimates across
CONUS from the NOAA operational HRRR model, which utilizes the RUC land
surface model (RUC LSM, Smirnova et al., 1997). We also include the NOAA
Climate Prediction Center (CPC) leaky-bucket hydrological model (Huang
et al., 1996; van den Dool et al., 2003) and in situ observations from
two nationwide networks: the NOAA/NCEI United States (US) Climate
Reference Network (USCRN; Bell et al., 2013) and the US Department of
Agriculture Soil Climate Analysis Network (SCAN; Schaefer et al., 2007).
This work provides an assessment of the similarities and differences of
soil moisture amounts and variance across three different products,
which are all used in various operational and research applications.
2 Soil Moisture Data
2.1 In Situ Observations
The study uses in situ soil moisture observations from two nationwide
networks. USCRN provides climate monitoring measurements of atmospheric
and soil properties. To increase the coverage of the in situ
observations, SCAN is also included. SCAN uses similar sensors to USCRN
(i.e., Hydra Probe sensors) and typically have volumetric soil moisture
(VSM; mwater3 /
msoil3) measurements at the same soil
depths (~5, ~10, ~20,
~50, and ~100 cm) as USCRN. Only data
from these five levels are used for consistency. Dirmeyer et al. (2016)
also found that these two networks have similar error variances. The
observations represent point measurements of the soil moisture at
specific sites across the US. Daily data are used in this study and
represent the average VSM of the entire 24-hour period based on local
standard time.
2.2 HRRR Model
The HRRR model is NOAA’s operational, convection-allowing model, which
has 3 km horizontal grid spacing and covers CONUS with a one-hour
temporal refresh rate (Dowell et al., 2022; James et al., 2022). In this
study, we use HRRRv3, which was operational between 12 July 12 2018 and
2 December 2020. The HRRR model utilizes a one-dimensional land surface
model (RUC LSM; Smirnova et al., 1997), which predicts heat and moisture
transfer vertically throughout the soil column. The RUC LSM has
undergone several enhancements over the years, including increasing its
resolution and incorporating new features, such as snow and ice models
(Smirnova et al., 2000; 2016). The current version predicts VSM at nine
vertical levels (0, 1, 4, 10, 30, 60, 100, 160 and 300 cm) and utilizes
cycling of soil conditions over several years to better capture the soil
moisture state. The HRRR utilizes moderately coupled land data
assimilation, meaning that near-surface atmospheric data assimilation
increments are used to adjust the soil analysis (e.g., Benjamin et al.,
2022). Given the recent and continued development of the HRRR data
assimilation system and land surface model, it is critical for
assessments of HRRR’s soil moisture to other estimates. This study
provides a benchmark for HRRR soil moisture estimates in support of the
continued development of NOAA’s land surface prediction capabilities in
the Unified Forecast System. The focus of this study is on the analyzed
soil moisture field, rather than forecast fields, and thus our results
are most directly relevant to the LSM and data assimilation system
development.
2.3 CPC Leaky-Bucket Model
The CPC soil moisture product utilizes a leaky-bucket model that solves
the time tendency equation in soil moisture over a region from several
inputs: precipitation minus evapotranspiration, net streamflow
divergence and net groundwater loss (Huang et al., 1996; van den Dool et
al., 2003). These inputs to the time tendency equation for soil moisture
have been improved over the years with new observations and
parameterizations (Fan and van den Dool, 2004; Arevalo et al., 2021).
The CPC model provides 1.6 m deep integrated soil moisture (ISM, mm),
and these estimates are provided daily for each of the NOAA climate
divisions across the United States (Guttman and Quayle, 1996). There are
typically about 7-10 climate divisions per state, although there are
fewer for states with smaller geographical areas, like those in the
Northeast United States. The CPC soil moisture data are used as an input
to the United States Drought Monitor (Svoboda et al., 2002) and
continues to be used as a reference data set in various soil moisture
application studies, from assessing soil moisture impacts on carbon
fluxes (e.g., Yao et al., 2021) to understanding climate impacts on
agricultural production (e.g., Atiah et al., 2022).
3 Methods
3.1 ISM and VSM Comparisons
Since the CPC product only provides 1.6 m ISM, VSM values from the in
situ and HRRR data are vertically integrated in order to compare 1.6 m
ISM in all three datasets. The VSM values are assumed to represent the
mean value over a depth between the midpoints of the specified levels,
as has been done in other studies (e.g., Dirmeyer et al., 2016; Ford and
Quiring, 2019). For example, the 10 cm VSM observation in the in situ
data is assumed to represent the average VSM for the layer between 7.5
cm (i.e., the midpoint between 5 and 10 cm) and 15 cm (i.e., the
midpoint between 10 and 20 cm). The 100 cm VSM in the in situ data is
also assumed to be constant to the depth of 160 cm. The HRRR ISM
calculations are better constrained than the in situ ISM calculations,
since the HRRR VSM data span 3.0 m below ground using 9 levels. VSM
values are also compared between the HRRR and in situ data to glean
whether certain vertical levels are driving differences between these
two datasets. An understanding of soil moisture differences at varying
depths is also critical since soil moisture’s role in Earth system
processes is depth dependent. In terms of spatial comparisons, the HRRR
and CPC data are linearly interpolated to the locations of the in situ
stations. The analysis is completed over a ~2.4 year
period from 12 July 2018 through 2 December 2020, which is the timeframe
that HRRRv3 was operational. By confining the analyses to this time
frame, uncertainties associated with model version changes are avoided.
3.2 Quality Control
In situ data provide the most direct physical estimate of soil moisture,
but it is important to ensure that the in situ data are of the highest
quality. As such, a variety of quality control procedures are
undertaken. First, in situ data are only included if they have VSM
values available at all five vertical levels (~5cm,
~10cm, ~20cm, ~50cm, and
~100cm), since missing data could lead to larger
uncertainties in the ISM calculation. Second, in situ stations directly
along the coast are removed due to unphysical spatial interpolations
from the CPC and HRRR data. From the remaining in situ data, we estimate
the ratio of error variance to ISM variance using the method defined in
Robock et al. (1995) and used in more recent soil moisture comparisons
studies (e.g., Dirmeyer et al., 2016). In essence, soil moisture can be
well approximated by a red-noise process (i.e., first-order Markov
process; e.g., Delworth and Manabe, 1988; Vinnikov and Yeserkepova,
1991) with the natural logarithm of the soil moisture autocorrelation
(r) decreasing linearly with increased lag times (𝜏). For this study,
autocorrelations are computed for 𝜏 of 1-30 days for each station’s ISM
daily anomalies. A linear fit is applied to the ln(r) versus 𝜏 data and
is extrapolated to 𝜏 = 0. Deviations from 1 at 𝜏 = 0 can be used to
solve for to the ratio of error variance to ISM variance (e.g., Robock
et al., 1995). For this study, stations where this ratio is greater than
0.08 are removed. This error ratio threshold is in-line with estimates
of the mean error ratio of the USCRN and SCAN networks (Dirmeyer et al.,
2016). While this error variance ratio threshold results in the removal
of 63 (~27%) of the 235 in situ stations, it provides
more confidence that only the highest quality in situ observations are
being used in the analyses. Even with the significant reduction of the
in situ data, the stations span the entirety of the CONUS (Figure 1,
dots). Different thresholds, autocorrelation lengths and dataset lengths
were tested and did not qualitatively impact the results.