The in situ observations generally have the highest ISM variance for most quintiles followed by CPC and HRRR, which has the lowest variances of the three datasets regardless of the wetness regime (Figure 3a-c). The in situ and CPC quintile mean ISM standard deviations are not significantly different for all quintiles except L20-40, while the HRRR quintile mean differences are significant for all quintiles when compared to both other datasets. Dirmeyer et al. (2016) showed that spatial scaling differences do not have a large impact (~10%) on in situ observation standard deviations via conducting tests where many stations that are separated by several km to up to 100 km are averaged together. Therefore, the differences between the in situ and NOAA modeled standard deviations are likely due to other factors outside of dataset spatial scale differences (e.g., model processes representation or model input data). Furthermore, the large variability or scatter in the ISM standard deviation differences between the in situ data and both models (Figure 3a-b) suggests that the cause of these differences depends on the specific location. The differences between HRRR and CPC ISM standard deviations, however, demonstrate a more systematic bias between these two modeling frameworks (Figure 3c).
Figure 3. Same as Figure 2, except for comparisons of ISM standard deviations (a-c) and VSM standard deviations (d-g).
5.2 VSM Standard Deviation Comparisons at Varying Depths
VSM standard deviations at four different depths are compared between HRRR and in situ data (Figure 3d-g) to determine whether a certain depth is driving the differences in the ISM standard deviations (Figure 3a). Regardless of the soil moisture regime, the HRRR surface VSM standard deviations compare well to the near-surface in situ observations. When averaging over all locations, the mean near-surface percentage differences in HRRR soil moisture from the in situ value is only +4.2% (Figure 3d). This is likely related to improvements made in the HRRR’s RUC LSM and the moderately coupled land data assimilation system that have been applied (Benjamin et al., 2022). However, at depths of 5 cm below ground and deeper (Figure 3e-g), most quintiles have statistically significant differences in the standard deviations. The mean percentage differences over all locations are -38.7%, -36.4% and -45.2% for the 5 cm, 10 cm and 100 cm levels, respectively, with the HRRR always having lower standard deviations than the in situ datasets for every quintile. These differences in VSM standard deviations are larger for wetter soil moisture regimes (L20-100). While the differences for the different soil moisture regimes are generally similar for the 5 cm, 10 cm and 100 cm depths, there are more extreme differences for specific locations at the 100 cm level. To summarize, the lower standard deviations in the HRRR ISM are being driven by the lower VSM standard deviations occurring below the surface level. It is important to note that both ISM and VSM standard deviation differences vary throughout the year, and the evolution of these differences as a function of month and season are provided in the supporting information document.
6 Conclusions and Future Work
A comparison of 1.6 m ISM between three different NOAA soil moisture products is conducted. This analysis uniquely includes the HRRR model with its RUC LSM, the CPC leaky-bucket model and in situ observations from two national networks. These soil moisture estimates are used in many operational and research applications, including atmospheric forecasting, drought monitoring, and assessing flood and fire risks. Therefore, quantifying differences in these NOAA models to observational networks across CONUS is critical.
Several conclusions are drawn from these comparisons.
1) The HRRR and CPC ISMs are both larger (i.e., wetter) in the driest regions and smaller (i.e., drier) in the wettest regions as compared to in situ observations.
2) These differences in the HRRR and in situ ISM amounts are largely caused by deep soil levels (~100 cm below ground). Shallower layers have similar trends to the deeper layers but have smaller differences, and thus a weaker contribution to the ISM differences.
3) The in situ observations have the largest ISM standard deviations, followed by the CPC leaky-bucket model and the HRRR model.
4) The HRRR soil moisture standard deviations compare well with the in situ standard deviations near the surface, but large differences are present at 5 cm below the surface and deeper.
The soil moisture differences presented in this study can be caused by a variety of reasons. In terms of modeled soil moisture, biases in the input datasets (i.e., precipitation or radiation), whether they come from a coupled atmospheric model in the case of HRRR or external sources in the case of CPC, have been shown to lead to biases in land surface model calculations (e.g., Mitchell et al., 2004; Min et al., 2021). Choices in the land surface model structure, such as the number and thickness of soil layers, the representation of soil and vegetation, and other model parameters, can also lead to biases in soil moisture prediction (e.g., Mitchel et al., 2004; Xia et al., 2014; 2015b). Min et al. (2021) found that snowmelt, freezing/thawing, and/or biases in precipitation and evapotranspiration led to differences in HRRR soil moisture as compared to in situ observations in New York and that the most relevant processes causing these differences varied throughout the year. The results in this study demonstrate consistent, region-dependent biases in NOAA modeled soil moisture as compared to in situ observations across CONUS, and future research should focus on understanding the model processes that are causing these biases.
Our results also provide important context to the current users of these models and observations. For example, HRRR’s land data assimilation system has recently undergone changes that primarily impact the near-surface soil state (Benjamin et al., 2022). The comparisons presented in this study do show better performance of HRRR soil moisture near the surface and thus may provide a first step towards understanding the impact of these model changes. Furthermore, these results can assist with the continued development and refinement of soil moisture models and products. The analyses presented here are currently being utilized for preparing training and validation data for a machine learning algorithm that uses data from the Advanced Baseline Imager on-board NOAA’s Geostationary Operational Environmental Satellite to estimate the soil moisture state at very high resolution (i.e., on the order of ~1 km). With a recent focus on land-atmosphere coupling and a continued shift towards higher-resolution models, such a product could be used as a supplementary input for strongly coupled land atmosphere data assimilation in the next generation of atmospheric models.
Acknowledgments
This work was supported by the NOAA FY21 High Performance Computing and Communications Program’s Information Technology Incubator. We would also like to acknowledge helpful feedback on and interest in this work from Liaofan Lin, Tanya Smirnova, Stan Benjamin, Curtis Alexander and Eric James.
Open Research
Several in situ and model datasets are used in this study. The USCRN data (Palecki et al., 2013) and the SCAN data (SCAN, 2016) will be archived upon publication. This archiving process is underway and will be with the Mountain Scholar repository through Colorado State University. We have uploaded a copy of these data as Supporting Information for the review process. The CPC data was accessed via https://ftp.cpc.ncep.noaa.gov/wd51yf/us/w_daily/ through the U.S. Data download link on the NOAA CPC product webpage (https://www.cpc.ncep.noaa.gov/products/Soilmst_Monitoring/). The HRRR operational model data (HRRRv3) was stored and accessed via the NOAA Hera supercomputer and is publicly archived at either https://registry.opendata.aws/noaa-hrrr-pds/ or https://console.cloud.google.com/marketplace/product/noaa-public/hrrr. The analysis code used to generate the analyses and figures in this manuscript are available at https://github.com/pjmarinescu/CIRA_Soil_Moisture and will also be archived in the same Mountain Scholar repository as the data upon publication.