SMAP Level-2 radar soil moisture data (SPL2SMA) are derived from SMAP High-Resolution Radar Sigma Nought, Version 3 (SPL1CS0) data.

SMAP Level-2 active soil moisture data are derived from SMAP High-Resolution Radar Sigma Nought, Version 3 (SPL1CS0) data. Utilizing the soil moisture algorithms discussed below, SMAP sigma0 measurements are converted into an estimate of the surface soil moisture of the 0-5 cm surface soil moisture in units of cm³/cm³ .

The following information has been adapted from Kim et al. 2014.

Soil Moisture Algorithms

The baseline retrieval algorithm inverts a forward scattering model. The equation for forward scattering (Eq. 2-1 in the ATBD) presents four unknowns: RMS height (s), dielectric constant (ε), correlation length (l), and vegetation water content (VWC). It will be shown that mv retrievals are insensitive to errors in knowledge of the correlation length over a wide range of m _v, roughness, and correlation lengths (Appendix 5.2) (Kim et al. 2012b). An exponential function is known to describe empirical measurements well (Mattia et al. 1997; Shi et al. 1997), which was used in the forward model and was not considered as an unknown during the retrieval. As an introduction to the discussion of the retrieval algorithm, and to place the discussion of each algorithm component in the proper context, the overall algorithm processing flow is described in the following section. Each portion of the algorithm flow, and the details of the retrieval methods and options, will then be described in detail in the subsequent sections.

Algorithm Flow

The algorithm flow of the SPL2SMA is presented in Figure 11 of the ATBD. The portion of the flowchart through the land surface classification is the initial processing done as a precursor to the actual retrievals. The SPL2SMA processor reads in 1 km resolution sigma0 from the SMAP Level-1C sigma nought product. The 1 km data in natural unit are aggregated onto the 3 km EASE-Grid, during which various quality flags are applied. Three quality flags are derived using the 3 km sigma0 freeze-thaw state (refer to the SPL3FTA ATBD), radar vegetation index (Section 3.4.2 of the SPL3FTA ATBD), and transient water body (Section 3.4.3 of the SPL3FTA ATBD). Static and dynamic ancillary data are sampled for each pixel of the SPL2SMA product. Sigma0 values from the past time are sampled and used by the time-series algorithm. For each 3 km pixel, land cover class information is obtained from the mostly annual ancillary data. The land cover information allows to choose an appropriate data cube for each pixel. The retrieval over the different land cover classes is spatially assembled to create a half-orbit output, followed by the conversion from the dielectric constant to soil moisture.

Baseline Algorithm: Time-series Data-cube Approach

Time-series Retrieval

The SMAP radar provides three independent channels (HH, VV, and HV). HV-channel measurements are reserved for possible use in correcting vegetation effects. The remaining two co-pol measurements (HH and VV) are not always sufficient to determine s and ε (see Section 3.2.4 in the ATBD) (Kim et al. 2012b). One of the main causes is the ambiguity in bare surface scattering: a wet and smooth surface may have the same backscatter as a dry and moderately rough surface has. Very often the time-scale of the change in s is longer than that of ε (Jackson et al. 1997). Then s may be constrained to be a constant in time, thus resolving the ambiguity (Kim et al. 2012b). The concept of ε time-invariant s has also been utilized in other studies (Joseph et al. 2008; Mattia et al. 2009; Verhoest et al. 2007). The SMAP baseline algorithm differs from these studies in that no ancillary or ground measurements or statistical assumptions are required to constrain s.

The SMAP baseline approach (Kim et al. 2012a ; Kim et al. 2012b) is a multichannel retrieval algorithm that searches for a soil moisture solution such that the difference between computed and observed backscatter is minimized in the least squares sense. The algorithm estimates s first and then retrieves ε_r using the estimated s. Vegetation effects are included by selecting the forward model's sigma0 at the VWC level given by an ancillary source or the SMAP HV measurements. The algorithm retrieves s and real part of dielectric constant (ε_r) using a time series of N co-pol backscatter measurements: σ⁰ HH ( t1), σ⁰ VV (t1), σ⁰ HH (t2), σ⁰ VV (t2),..., σ⁰ HH (t N), and σ⁰ VV (tN).

There are thus 2 N independent input observations and N+1 unknowns consisting of N ε_r values and one s value. Note that the VWC provided by ancillary information is allowed to be varying throughout the time series.

Radar backscattering coefficients before conversion to decibels can be modeled as Gaussian random variables (Ulaby et al. 1986) to account for speckle and thermal noise effects. Assuming sufficient integration following power detection, the backscattering coefficient after conversion to decibels can also be modeled as a Gaussian random variable. Because SMAP will observe HH and VV returns at slightly different center frequencies, the effects of speckle and thermal noise on these measurements are statistically independent. Statistical independence of speckle in measurements at differing time steps is also expected. These facts and a maximum likelihood formulation motivate least-squares retrieval approaches based on the average of individual error terms. It is noted that calibration, radio frequency interference, and other error sources may produce correlated error terms. The systematic and correlated components from these sources will be corrected. Any residual s may impact overall algorithm performance and are modeled as uncorrelated Gaussian noise in the  ATBD for this product (although they may still contain correlated noise).

Additional details are provided in the ATBD.

Snapshot Retrieval

When the number of time-series input is set to 1, the time-series method becomes a conventional snapshot approach. ε_r and s are adjusted simultaneously at each time instance so as to minimize the distance between the data cube prediction and the observation. The snapshot retrieval algorithm was also tested with the two sets of field campaign data used in Section 3.2.3 of the  ATBD. Figure 14a in the ATBD shows that the bare surface retrieval has an RMSE of 0.055 cm³/cm³, which is larger than the error of the time-series retrieval. Closer examination however indicates that the retrieval e rror increases for larger soil moisture. When the performance of the Monte Carlo simulation of the Kp effect is compared between time-series (Figure 12a of the ATBD) and snapshot (Figure 12c of the ATBD) retrievals, the performance difference is much greater than the in situ validation shows. The radar measurement error of the in situ data (~0.3 dB) is smaller than simulated by the Monte Carlo analysis, which may explain why the Monte Carlo analysis of the snapshot retrieval show the poorer performance.

The snapshot retrieval over the pasture surface is poor. Figure 14b in the ATBD shows a retrieval error of 0.105 cm³/cm³. It is possible that the structure of the grass was important in characterizing the backscatter to be very different from that of the bare surface. The possibility is supported also by the large retrieval error (not shown) when the retrieval was performed using the bare surface data cube (NMM3D) after treating the grass surface as if it is a bare surface. The evaluation results suggest the limited applicability of the snapshot retrieval, and highlights the need for time-series retrievals.

Optional Algorithm

Change Detection by Kim and van Zyl (KVZ)

The full details of this algorithm are described by Kim and van Zyl (2009). A brief summary is presented here. An expression will be derived to relate the backscattering cross section to soil moisture as:

m_v = f(σ_hh, σ_vv)

where the normalized radar cross sections are expressed in decibels. Since this expression depends on the biomass level, the cross-polarization will be used to compensate the biomass variation over time: how this will be implemented is being studied . One potential choice for f is given by:

  (Equation 1)

These two coefficients for each pixel will be determined using the expected minimum and maximum values for soil moisture and the time-series backscattering cross section data. Accommodating different vegetation phenology during the implementation of the above relationship remains as a task. The obvious assumption contained in Equation 1 is that there is a linear relationship between the radar cross-sections and the soil moisture. However, simulated data suggests that this relationship may be nonlinear. Ground radar and airborne data will be used to derive the final expression for the function f(σ_hh, σ_vv).

The slowly varying backscattering cross section component due to the biomass variation will be estimated and compensated using the cross-polarization backscattering cross section. Ground radar measurements will help derive the expres sion before the SMAP instrument is launched. However, since the g round measurements are limited, with simulated data various vegetation environments will be considered. When the two extreme conditions (completely dry and wet) of the soil moisture and radar backscatter are known exactly, a test with a corn field experiment shows that the soil moisture may be retrieved with an error of 0.026 cm³/cm³ (Table 3 of Kim and van Zyl). When the errors in the two extreme values are simulated with the random noise, the retrieval error reaches 0.05 cm³/cm³. Alternatively, Kim and van Zyl proposed that C₁ at each radar pixel (e.g., at 3 km resolution) is inverted using the m_v retrieved at a coarse scale (such as at 36 km resolution) using the radiometer records. The radar-radiometer method is the subject of the  ATBD for the SPL2SMAP and SPL3SMAP products.

Additional details are provided in the ATBD.

Note: Algorithms that are currently not being used for these data include Snapshot, Dubois/van Zyl (DVZ), and Shi. Thus, the following data fields contain a fill value of -9999.0:

• soil_moisture_snapshot
• soil_moisture_snapshot_DVZ
• and soil_moisture_snapshot_shi

Algorithm Inputs and Outputs

Table 3 lists the required input data sets to the radar-only soil moisture inversion. These include the primary choices that meet the ancillary data requirements. Ancillary data sets are defined as external data sets that are required as inputs to SMAP retrieval algorithms in the generation of the standard Level-2 and Level-3 active products. Ancillary data needed by the SMAP mission fall into two categories: static ancillary data, which do not change during the mission, and dynamic ancillary data, which require periodic updates in timeframes ranging from seasonally to daily. Static data include parameters such as permanent masks (land/water/forest/urban/mountain), the grid cell average elevation and slope derived from a Digital Elevation Model (DEM), permanent open water fraction, and soils information (primarily sand and clay fraction). All static ancillary data are resampled to the 3 km EASE-Grid 2.0 as the output products. The dynamic ancillary data include land cover, precipitation, and effective soil temperatures. The choice of which ancillary data set to use for a particular SMAP product is based on a number of factors, including its availability and ease of use, and its inherent error and resulting impact on the overall soil moisture or freeze/thaw retrieval accuracy. Latency, spatial resolution, temporal resolution, and global coverage are also important. The choice of a primary source for each of the ancillary data parameters is fully documented in individual SMAP Ancillary Data Reports (refer to Section 6.1 in the ATBD). Brief descriptions are provided below for the input radar backscatter, and for the ancillary parameters generated by the SPL2SMA processor.

Radar Backscatter

As previously stated, the primary input to the SPL2SMA algorithm is the Level-1C gridded radar product. This product consists of the calibrated high-resolution (SAR-mode) multiple polarization radar backscatter measurements and associated parameters, Earth-located and mapped to a swath-based, along-track/cross-track, 1 km EASE-Grid 2.0. It is assumed that the backscatter measurements have been fully calibrated and corrected for atmospheric attenuation and ionospheric effects such as Faraday rotation. Information provided in the Level-1C data set or in ancillary calibration files should enable estimates of the sigma nought (σ0) accuracy (Kp) for each measurement based on instrument characteristics.

Vegetation Water Content derived from Radar Vegetation Index

The SPL2SMA algorithm requires information on the VWC. One potential source of vegetation information is a polarimetric parameter derived from SMAP data such as the Radar Vegetation Index (RVI) (Kim and van Zyl 2000).

Alternatively, the vegetation amount may be derived based on low-order polynomial of the NDVI and optical remote-sensing images (refer to the SPL2SMP ATBD). In principle the RVI or any other parameter based on the radar backscatter should correlate with the vegetation amount more strongly than the optical NDVI, because fundamentally optical radiation cannot penetrate vegetation canopies while the L-band microwave signal can. Furthermore the RVI is concurrent with the radar observation. Studies are ongoing with regard to whether the radar backscatter may improve estimates of vegetation effects. The Data Fields Ancillary Data group contains VWC derived from both RVI and NDVI. 

Radar Water Bodies

An algorithm for discriminating transient water bodies from land surfaces is also part of the SPL2SMA processor. There are many combinations of water surfaces adjoined by land surfaces, (depending on both land and water surface roughness and presence of vegetation) that affect this discrimination capability. At present, the case of a calm water surface next to lightly vegetated land was studied. It was found that a threshold (3 dB) applied to the ratio of HH to VV effectively separated water from land at SMAP's 40 deg incidence angle. In this study, Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) images were analyzed over fairly flat terrain. Despite the apparent differences between UAVSAR and SMAP, the measurement quality for both instruments is determined by the same statistical processes inherent in SAR imaging. Calibration error and the number of single looks of one UAVSAR pixel are similar to those of one pixel of the SMAP SAR.The algorithm was tested over numerous locations with a detection error of smaller than 10% (Kim et al. 2014).

Refer to the Data Fields document for a description of all data fields.

Anthropogenic Radio Frequency Interference (RFI), principally from ground-based surveillance radars, can contaminate both radar and radiometer measurements at L-band. Early measurements and results from ESA's Soil Moisture and Ocean Salinity (SMOS) mission indicate that in some regions RFI is present and detectable. The SMAP radar and radiometer electronics and algorithms include design features to mitigate the effects of RFI. The SMAP radar utilizes selective filters and an adjustable carrier frequency to tune to predetermined RFI-free portions of the spectrum while on orbit.

More information about error sources is provided in Section 3.5: Error Budget of the ATBD.