Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Characterization of vegetation and soil scattering mechanisms acrossdifferent biomes using P-band SAR polarimetry

Seyed Hamed Alemohammada,b,*, Alexandra G. Koningsc, Thomas Jagdhuberd,Mahta Moghaddame, Dara Entekhabia

a Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, United Statesb Department of Earth and Environmental Engineering, Columbia University, United Statesc Department of Earth System Science, Stanford University, United Statesd Microwaves and Radar Institute, German Aerospace Center (DLR), Germanye Department of Electrical Engineering, University of Southern California, United States

A R T I C L E I N F O

Keywords:AirMOSSP-bandPolarimetric decompositionSARSoil moistureVegetation

A B S T R A C T

Understanding the scattering mechanisms from the ground surface in the presence of different vegetationdensities is necessary for the interpretation of P-band Synthetic Aperture Radar (SAR) observations and for thedesign of geophysical retrieval algorithms. In this study, a quantitative analysis of vegetation and soil scatteringmechanisms estimated from the observations of an airborne P-band SAR instrument across nine different biomesin North America is presented. The goal is to apply a hybrid (model- and eigen-based) three component de-composition approach to separate the contributions of surface, double-bounce and vegetation volume scatteringacross a wide range of biome conditions. The decomposition makes no prior assumptions about vegetationstructure. We characterize the dynamics of the decomposition across different North American biomes and assesstheir characteristic range. Impacts of vegetation cover seasonality and soil surface roughness on the contribu-tions of each scattering mechanism are also investigated. Observations used here are part of the NASA AirborneMicrowave Observatory of Subcanopy and Subsurface (AirMOSS) mission and data have been collected between2013 and 2015.

1. Introduction

Soil moisture is the key state variable that controls the terrestrialwater, carbon, and energy fluxes between land surface and atmosphericboundary layer, mainly by regulating photosynthesis and surface eva-poration (Seneviratne et al., 2010). By constraining the partitioning ofenergy fluxes between latent and sensible heat fluxes in water-limitedregions, soil moisture also plays a significant role in the prediction skillof weather and climate models (Entekhabi et al., 1996; Koster et al.,2010). Soil moisture has a memory that captures the anomalies inprecipitation and radiation and can be used to identify regions of strongfeedback between land surface and atmospheric boundary layer(McColl et al., 2017). Moreover, vegetation stress, and subsequentlyphotosynthetic activity, depends on the amount of water availablethrough the roots (as well as atmospheric conditions). Therefore,knowledge of Root Zone Soil Moisture (RZSM) (and where applicable,interaction of roots, soil moisture, and the water table) are necessary toaccurately model evapotranspiration seasonality (Thompson et al.,2011). Characterization of the spatio-temporal patterns of soil moisture

with depth, therefore, enables improved predictions of the response ofplants to the changing climate. Thus, remotely sensed, large-scale es-timates of root-zone soil moisture have a number of operational andscientific use in hydrometeorology and ecology, if they can be obtained.

The penetration depth associated with microwave remote sensingsoil moisture increases as the electromagnetic frequency of the mea-surement decreases (Ulaby et al., 2014). Current global microwavesatellite observations of soil moisture are limited to those at L-bandfrequency and higher due to spectrum availability, readiness of sciencealgorithms and technological restrictions of obtaining reasonablespatio-temporal resolution from low-earth orbit satellites (Entekhabiet al., 2014; Kerr et al., 2010). However, L-band instruments, are sen-sitive to only a few centimeters of the top soil layer. Detection of RZSMrequires P-band instruments, which have a penetration depth of severaltens of centimeters, depending on soil texture, the profile of soilmoisture content and vegetation cover (Moghaddam et al., 2007;Konings et al., 2014). The future BIOMASS mission will carry a fullypolarimetric P-band Synthetic Aperture Radar (SAR) instrument andprovide an opportunity to estimate RZSM globally (except over North

https://doi.org/10.1016/j.rse.2018.02.032Received 8 August 2017; Received in revised form 30 January 2018; Accepted 20 February 2018

* Corresponding author at: Department of Civil and Environmental Engineering, Massachusetts Institute of Technology, 77 Massachusetts Ave., Cambridge, MA 02139, United States.E-mail address: hamed_al@mit.edu (S.H. Alemohammad).

Remote Sensing of Environment 209 (2018) 107–117

0034-4257/ © 2018 Elsevier Inc. All rights reserved.

T

America and Europe) from satellite-based SAR observations (Toanet al., 2011; Carreiras et al., 2017).

The Airborne Microwave Observatory of Subcanopy and Subsurface(AirMOSS) mission (a NASA Earth Venture-1 project) is the first P-bandairborne campaign to estimate RZSM and use that to better characterizeNet Ecosystem Exchange (NEE) across North America (Allen et al.,2010). AirMOSS employs an airborne fully polarimetric SAR operatingat P-band (430MHz) to monitor the dynamics of RZSM across ten sitesin North America, covering nine different biomes representative of theentire North American continent (Table 1). The biomes across AirMOSSsites range from tropical and temperate forest to boreal transitionalforest and evergreen needle-leaf as well as cropland, woody savanna,grassland and shrubland. AirMOSS backscatter estimates have a highradiometric calibration accuracy of 0.5 dB (Chapin et al., 2015), and anoise equivalent σ° of −40 dB (Tabatabaeenejad et al., 2015) whichmakes them suitable for accurate soil and vegetation parameter esti-mation (although the data require a phase calibration for polarimetricapplications, as further discussed below).

AirMOSS successfully conducted four years (2012–2015) of air-borne campaigns, with 186 science flights each covering an area ofapproximately 100 km×25 km. Each site was visited for 2 or 3 cam-paigns per year. For each campaign, 2 to 3 overflights were performed afew days apart. Data from these campaigns provide an opportunity todevelop and validate new algorithms for RZSM, vegetation and surfaceproperties retrieval and to analyze and quantify the contributions ofvolume and surface scattering, and their interactions to P-band signalsacross a wide range of vegetation types. In this study, we provide aquantitative analysis of the variability of surface and vegetation para-meters across the campaign sites. The AirMOSS radar calibration modelwas updated in April 2013, so we restrict our study to observationsfrom the period April 2013 through September 2015 to ensure a con-sistent calibration accuracy/quality.

Unlike at L-band, the variability of the soil moisture with depth issignificant over P-band penetration ranges and must be accounted for,which means calculating only a single equivalent soil moisture valuemay result in values that do not represent the average profile due to theeffects of subsurface reflections (Konings et al., 2014). Hence, twodifferent approaches have previously been used for retrieval of soilmoisture profiles from AirMOSS observations using modeled vegetationand ancillary vegetation parameters. For campaign sites with mono-species woody or non-woody vegetation (including mono-speciesforested sites), the approach is to model the vegetation using a detailedscattering model that consists of a stem layer and a canopy layer(Burgin et al., 2011), and assume a second order polynomial shape forthe soil moisture profile. Parameters of the vegetation volume scat-tering model are derived based on field measurements, and the coeffi-cients of the polynomial profile are retrieved using snapshot measure-ments (Tabatabaeenejad et al., 2015). This approach also assumes a

temporally and spatially constant value for the surface roughnessparameter, determined by site-specific calibration, for each land covertype to reduce the number of unknowns in the retrieval.

A second retrieval algorithm using AirMOSS observations is focusedon campaign sites with multi-species woody vegetation cover (Truong-Loi et al., 2015). This approach uses a semiempirical inversion model toestimate soil moisture profile, surface roughness and the abovegroundbiomass. The parameters of the scattering model are estimated by re-gressing the semiempirical model to forward estimates of the fullscattering model using field measurements from the Forest InventoryAnalysis (O’Connell et al., 2013). In order to increase the number ofobservations and make the system of equations well-defined, a timeseries scheme is designed that assumes constant surface roughness andabove ground biomass across the three flight days of each AirMOSScampaign (2 to 3 observations over 7–10 days). RZSM retrievals fromboth of these algorithms meet the AirMOSS mission requirements of anunbiased Root Mean Squared Error (ubRMSE) of 0.05m3/m3 for soilmoisture over the AirMOSS sites when validated against ground mea-surements (Tabatabaeenejad et al., 2015; Truong-Loi et al., 2015).

Using vegetation (and roughness) parameters that are dependent onsite-specific field measurements limits the applicability of thesemethods outside of the United States (where FIA data are not available)and in areas where in situ measurements may be logistically difficultand expensive. Furthermore, these approaches limit the applicability ofthe retrieval algorithm across diverse and large-scale land covers andmay create errors due to variability in vegetation structure even acrossa single site. Hence, the transferability of the approach to other test sitesor remote regions is hardly given. However, the design of retrieval al-gorithms that require fewer parameters is complicated by the limitedamount of information contained in the backscattering coefficients. Useof the coherent fully polarimetric observations, which was not im-plemented in the existing AirMOSS retrieval algorithms, provides ameans to overcome this potential problem.

A quantitative understanding of scattering mechanisms from landsurfaces with vegetation cover is required for interpreting P-band SARobservations. Understanding the relative role of attenuated groundbackscatter, vegetation volume backscatter and interactions term basedon the SAR observations serves as a guide to the design of parsimoniousRZSM retrieval algorithms. The goal of this study is to quantify thecontributions of ground (surface and double-bounce) and volumescattering across the wide range of vegetation covers in fully polari-metric P-band SAR observations. Such an extensive analysis was notconducted before due to the lack of fully polarimetric P-band data withreasonable SNR over different land covers/biomes and across differentseasons.

We use AirMOSS coherent fully polarimetric observations (phaseand amplitude information). To begin, it is necessary to overcome thelack of a dedicated phase calibration in the AirMOSS observations.First, we preprocess the observations by merging the four flight lines ineach campaign, removing pixels with high topographic slope and cali-brating the polarimetric phase of the observations. Then, we apply afully polarimetric decomposition model to estimate the contribution ofeach scattering mechanism to the total backscattering power.Vegetation scattering is modeled using a cloud of randomly-orienteddipoles, without the need for prior assumptions on vegetation para-meters. Applying the estimation approach to observations across allAirMOSS campaign sites, we characterize the temporal and spatialdifferences in the relative contributions of the different scattering me-chanisms. These results can provide guidance for the design of futurelow-frequency RZSM retrieval algorithm.

The rest of the study is organized as following: Section 2 reviews thedata pre-processing and presents estimates of phase bias from AirMOSSobservations. Section 3 describes the fully polarimetric model used todecompose the scattering mechanisms, and Section 4 outlines the es-timation steps. Section 5 presents the results, and conclusions areprovided in Section 6.

Table 1List of AirMOSS campaign sites and their biome type.

Name & location Biome type

BERMS, Saskatchewan, Canada Boreal forest/evergreen needle-leaf, mixedforest, cropland

Howland Forest, ME, USA Boreal transitional/mixed forestHarvard Forest, MA, USA Boreal transitional/mixed forestDuke Forest, NC, USA Temperate forest/mixed forest, croplandMetolius, OR, USA Temperate forest/evergreen needle-leafMOISST, Marena, OK, USA Temperate grasslands/cropsTonzi Ranch, CA, USA Mediterranean forest/woody savannaWalnut Gulch, AZ, USA Desert and shrub/open shrubland and

grasslandChamela, Mexico Subtropical dry forest/broadleaf deciduous,

crops, woody savannaLa Selva, Costa Rica Tropical moist forest/evergreen broadleaf,

crops

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

108

2. Data preprocessing

The AirMOSS instrument has three independent polarization chan-nels (i.e. HH, VV, VH) measured with amplitude and phase. AirMOSSdata products provide the HH, VV and VH polarimetric channels andreciprocity (equivalence of VH and HV) cannot be tested with the dataas processed today. The incidence angle of the processed image pro-ducts ranges from 25° to 55° (Chapin et al., 2012). In our decompositionmodel, we assume the observations follow reflection symmetry(Section 3). AirMOSS measurements in each campaign site include fouradjacent flight lines with few overlapping pixels. For the purpose of thisstudy, we have merged the flight lines for each campaign and for theoverlapping pixels, measurements with an incidence angle closer to 40°are selected (chosen to be in the middle range of the incidence anglesfrom AirMOSS instrument). Moreover, pixels with high topographicslope are excluded from this analysis. Steep surfaces act like a rotatedsurface, and the assumption of reflection symmetry is violated(Freeman and Durden, 1998). Slopes are calculated using digital ele-vation data and provided by the science team as part of the ancillarydata in each campaign. We choose a threshold of 15% slope and re-moved pixels that have a higher slope in either North-South or East-West direction. Slopes in any other direction (e.g. North-west–Southeast) have a component in the North-South and East-Westdirections. Therefore, the two major directions are the optimal way toidentify regions of high slopes. This results in removal of between 5%and 45% of the pixels across different sites.

All the data from the AirMOSS mission are radiometrically andpolarimetrically calibrated by the project science team at JPL and re-ported in documentation files of each campaign site (available athttps://daac.ornl.gov/cgi-_bin/dataset_lister.pl?p=36) and in Chapinet al. (2012). The only aspect of the data which is not calibrated in theproduct release is the phase information in the HHVV channel, as pri-marily intensity-based parameter retrieval approaches were planned forapplication. Hence, no trihedral corner reflector was deployed in theflight lines. Therefore, in order to take advantage of the coherent po-larimetric information in the observations, we propose to use a knowntarget, i.e. bare flat land, to calibrate the S S *HH VV phase information.While this approach is observation-based and not ideal (totally baresurfaces may not be completely devoid of any vegetation), it is our onlypossible solution to calibrate the HHVV-phase information subsequentto acquisition.

Here, we use the distribution of measurements across low vegeta-tion pixels (as a canonical target) for phase calibration (similar to theapproach developed by Zebker and Lou, 1990). Unlike the approachproposed by Zebker and Lou (1990), we use cross-pol observations toidentify low vegetation pixels. For these pixels, the phase of the S S *HH VVin backscattering should ideally be zero (Guissard, 1994). Therefore, weuse the phase distribution of this signal across many bare soil/low ve-getation pixels to estimate the phase offset from zero, if any.

Static land cover maps cannot account for seasonally varyingchanges in whether a pixel has low vegetation cover, and underestimatethe number of low vegetation pixels. Instead, we use the data-drivencriteria of −40 dB< < −S S* 25HV HV dB to identify bare soil and lowvegetation pixels. This criterion is selected based on the fact that ve-getation cover leads to relatively high cross-polarized backscattering;therefore, a low cross-polarized backscatter value indicates low or novegetation coverage. Soil roughness can also contribute to the cross-polarized signal, but we expect that to be negligible for P-band and non-tilled soils. We find that the [−40,−25] dB range is adequate to isolatenearly bare surfaces. The −40 dB is the noise floor (σ°) for AirMOSSdata (Chapin et al., 2012; Hensley et al., 2015). Therefore, any scat-tering below that value is considered noise. The upper bound −25 dB isselected based on our evaluation of the distribution of phase informa-tion across all sites. Below this value, we do not see any trend in thephase value, which means the surface conditions are stable and there isno substantial vegetation scattering. To increase the accuracy of this

data-driven detection, we further excluded pixels that are classified aseither developed, open water or wetland based on the NLCD land coverdata.

We applied this approach to all the campaigns used in this study.Fig. 1 shows the distribution of S S *HH VV phases across bare soil/lowvegetation pixels for one of the campaigns at each site conducted in2015 (as an example of the variability of the phase bias across differentcampaign/sites). We use the median of phase across these pixels in eachflight campaign as the bias in the phase information, and correctS S *HH VV observation for that flight by removing the phase bias. There-fore, the phase bias varied from each campaign and site to another. Forthe 2015 flights shown Fig. 1, the estimated bias ranges from 82° to 95°.

3. Forward physical model of vegetated surface

A key challenge in monitoring soil moisture using radar measure-ments is that the interaction of signal with vegetation structure andwater content is complex for a general vegetation cover. Moreover, thebackscatter signal has contributions directly from the vegetation andbounced signal between the surface and vegetation (double-bounce)together with the backscatter from the surface. Since the vegetationcharacteristics are not known a priori, they add more unknowns to themoisture estimation problem. This problem is exacerbated when usinglow-frequency (e.g., P-band) observations to estimate RZSM. Soilmoisture can vary significantly across the root-zone, with the shape ofthe profile varying in time. Interactions and phase delays between re-flections from different soil layers cause the equivalent half-space soilmoisture values to differ significantly from the true average soilmoisture profile, so that soil moisture at multiple depths has to be re-trieved simultaneously (Konings et al., 2014).

Two main approaches have been developed to decompose fullypolarimetric SAR observations. First, eigen-based decomposition that isa mathematical data-driven technique to interpret eigenvalues and ei-genvectors of the coherency matrix as physical mechanisms con-tributing to the backscattering observation (Holm and Barnes, 1988;van Zyl, 1989; van Zyl, 1992; Cloude and Pottier, 1996). Second,model-based decomposition that develops a physical scattering modeland retrieves the parameters using the covariance or coherency matrixof observations (Freeman and Durden, 1993; Freeman and Durden,1998; Freeman, 2007; Yamaguchi et al., 2005). Eigen-based

Fig. 1. Distribution of S S *HH VV phase across different sites for one campaign in 2015. Thephase data from pixels that have a |SHV|2 in the range [−32− 26] dB are included in thisfigure. The central red line indicates the median, the edges of the box are 25th and 75thpercentiles, the whiskers show the most extreme values. (For interpretation of the re-ferences to color in this figure legend, the reader is referred to the web version of thisarticle.)

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

109

decompositions have no pre-assumption on the type of scatteringcomponents and result in orthogonal components while in model-basedapproaches scattering components need to be known in advance and donot necessarily lead to orthogonal components. It is not straightforwardto provide physical interpretation of eigen-based decompositions;however, scattering components in model-based approaches can bechosen beforehand and adapted to the scattering scenario (e.g. for P-band). In eigen-based decomposition, only rank-1 components can bedecomposed (vegetation is mostly a rank-3 component) while compo-nents with different ranks can be designed and combined within thedecomposition procedure in model-based approaches.

Nevertheless, both approaches are widely used in decomposing SARobservations and retrieving soil moisture and vegetation parameters(Moghaddam et al., 2000; Hajnsek et al., 2009; Kim and van Zyl, 2009;van Zyl et al., 2011; Arii et al., 2011; Jagdhuber et al., 2012; Jagdhuberet al., 2013; Kim et al., 2014; Jagdhuber et al., 2015; Alemohammadet al., 2016; He et al., 2016). Here, we use a comprise and combine themodel-based decomposition approach of Freeman and Durden (1998)with an eigen-based alpha-angle approach (as in Jagdhuber et al.,2015) to benefit from the capabilities of both techniques.

We use a three component scattering model and decompose thetotal backscattering signal to contributions from ground surface scat-tering, vegetation volume scattering and double-bounce scattering be-tween the vegetation and ground surface (Freeman and Durden, 1998,1993; Lee and Pottier, 2009; Yamaguchi et al., 2006):

= + +T T T Tobs S V D (1)

in which S, V, and D denote ground surface, volume, and double-bouncescattering mechanisms, respectively. ⟨⟩ denotes spatial ensembleaverage in the data processing. T represents the scattering coherencymatrix:

where superscript * denotes complex conjugate. For the model used inthis study, we assume reflection symmetry; therefore, in the Tmodel,T13= T23= 0. We analyze the observation covariance matrices andestimate the median values of each channel across all the flights foreach site. The HVVV and HHHV channels have median power that is15–20 dB lower than the co-pol and 10–15 dB lower than cross-polobservations, and marginal co-to-cross-polarization correlations thatare indicative of reflection symmetric scattering.

Vegetation is modeled assuming the vegetation layer consists ofcylindrical particles that are randomly oriented with a uniform prob-ability density function (PDF). The coherency matrix for vegetationvolume scattering is (Yamaguchi et al., 2005, 2006):

= ⎡

⎣⎢

⎤

⎦⎥T f

0.50 0 00 0.25 00 0 0.25

Vv

(3)

in which fv represents vegetation volume scattering intensity. For sur-face and double-bounce scattering we use the α scattering definition tofind their corresponding contribution (Cloude, 2010). The α scatteringangle is calculated using an eigen-based decomposition of the coher-ency matrix, and defined as the inverse cosine of the length of the firstelement of the first eigenvector of the coherency matrix Trem. A detailedderivation of the α scattering angle is presented in Cloude and Pottier(1997), Cloude et al. (2001) and Jagdhuber (2012). Assuming surfaceand double-bounce scattering components are orthogonal rank-1, thecoherency matrix for direct backscattering from the surface would be

(Jagdhuber et al., 2015; Cloude, 2010):

=⎡

⎣⎢⎢

−−

⎤

⎦⎥⎥

T fα αα α

cos sin 0cos sin 00 0 0

Ss

s s

s s

2

2

(4)

in which fs represents intensity of surface scattering. Double-bouncescattering is represented with (Jagdhuber et al., 2015; Cloude, 2010):

=⎡

⎣

⎢⎢

⎤

⎦

⎥⎥

T fα αα α

sin cos 0sin cos 0

0 0 0

Dd

d d

d d

2

2

(5)

in which fd represents the intensity of double-bounce scattering.The model presented here can capture the interactions of a low-fre-

quency SAR signal within the vegetated surface with relatively simplerepresentation of volume scattering. The focus of this study is on char-acterizing the contribution of each scattering mechanism; therefore, weestimate αs and αd parameters together with fs, fd, and fv rather than theexplicit surface and vegetation reflectivities (or dielectric properties) andsurface roughness. This analysis is aimed at estimating the three scatteringdecomposition across a wide range of vegetation covers (biomes) using P-band polarimetric observations alone (no ancillary information is used).

4. Estimation approach

In this section, we outline the estimation approach to quantify thecontribution of each scattering mechanism. We use a hybrid (model-and eigen-based) approach and use the physical model introduced inthe previous section along with an orthogonality criterion on the twoeigen-based α scattering angles to separate the surface and double-

bounce scattering contributions (Cloude, 2010; Jagdhuber et al., 2015).Observations from AirMOSS instrument are provided in terms of the

elements of the covariance matrix. Therefore, we use the special unitarytransformation matrix to transform the covariance matrix to coherencymatrix (Lee and Pottier, 2009; Cloude, 2010). The observed coherencymatrix after implementing the phase calibration is:

=⎡

⎣

⎢⎢⎢

⟨ + ⟩ ⟨ + − ⟩⟨ − + ⟩ ⟨ − ⟩

⟨ ⟩

⎤

⎦

⎥⎥⎥

T

S S S S S SS S S S S S

S

12

| | ( )( )* 0( )( )* | | 0

0 0 4 | |.obs

HHobs

VVobs

HHobs

VVobs

HHobs

VVobs

HHobs

VVobs

HHobs

VVobs

HHobs

VVobs

HVobs

2

2

2

(6)

The first step in the estimation process is to estimate fv and removethe vegetation contribution from the observations. Based on the modelpresented in the previous section, only volume scattering contributes tothe cross-pol channel; therefore, we can estimate fv directly from thecross-pol observation (T33). Previous studies have shown that the cross-polarized channel can have contributions from other scattering me-chanisms (ground surface and double-bounce) and this might result in ahigher value for fv that is physically impossible (van Zyl et al., 2011).Therefore, fv estimation should be bounded so that after removing thevegetation contribution the remained power in surface and double-bounce scattering is non-negative (a mathematical constraint). Thisimplies that the eigenvalues of the Trem coherency matrix should be

=⎡

⎣

⎢⎢⎢

⎤

⎦

⎥⎥⎥

=⎡

⎣

⎢⎢⎢

⟨ + ⟩ ⟨ + − ⟩ ⟨ + ⟩⟨ − + ⟩ ⟨ − ⟩ ⟨ − ⟩

⟨ + ⟩ ⟨ − ⟩ ⟨ ⟩

⎤

⎦

⎥⎥⎥

T

T T TT T TT T T

S S S S S S S S SS S S S S S S S S

S S S S S S S

** *

12

| | ( )( )* 2 ( ) *( )( )* | | 2 ( ) *

2 ( )* 2 ( )* 4 | |

HH VV HH VV HH VV HH VV HV

HH VV HH VV HH VV HH VV HV

HV HH VV HV HH VV HV

11 12 13

12 22 23

13 23 33

2

2

2(2)

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

110

non-negative (van Zyl et al., 2011):

= − ⎡

⎣⎢

⎤

⎦⎥T T f

0.50 0 00 0.25 00 0 0.25

.rem obsv

(7)

The objective function to estimate fv by incorporating the upperbound limit is:

=≥

f fT

max( )subject to eig( ) 0.

v vrem (8)

Here, the eigenvalues of Trem are numerically calculated for differentvalues of fv, and the maximum physically feasible fv is estimated.

Next, we estimate αs, αd, fs and fd using the equations provided inJagdhuber et al. (2015). These equations provide solutions for thesefour variables with an ambiguity for αs and αd which is resolved basedon an orthogonality condition = −( )α αs

πd2 , inherent in the physics of

the alpha scattering model (Cloude et al., 2001). Among the two esti-mated α′s, the one that is smaller than π

4is αs, and the other one is αd

(Cloude, 2010). Using this definition, neither of the surface or double-bounce scatterings has to be selected artificially to be the dominant one(Freeman and Durden, 1998) and the physics of the alpha scatteringmechanisms (orthogonality criterion) regulates the decomposition ofthe ground components. Hence, the respective scattering powers aredivided between them accordingly. In the next section, we present theresults of applying this estimation approach to all campaign sites of theAirMOSS mission.

5. Results

In this section, we present the results of applying the new decom-position approach to observations from all the campaign sites fromAirMOSS mission. These include 167 campaigns across all the 10 sitesintroduced in Table 1. First, we present statistics across all sites. Then,we discuss example cases from some sites and provide directions forfuture retrieval algorithm efforts.

5.1. Overall statistics

We use three relative indices to compare the contribution of thescattering mechanisms across different sites. The indices are definedbased on the span of the coherency matrix of each scattering compo-nent:

= += +=

= < + > + < > + < >

P f α αP f α αP f

P S S S S

( cos sin )( cos sin )

( | | 2 | | | | )

s s s s

d d d d

v v

T HH VV HV VV

2 2

2 2

12

2 2 2(9)

in which Ps, Pv, and Pd represent the span of the coherency matrix forsurface, vegetation, and double-bounce scattering, and PT is the totalspan of the observation coherency matrix. The three relative indices arePs/PT, Pv/PT, and Pd/PT which show the normalized contribution ofsurface, vegetation and double-bounce scattering, respectively. Fig. 2shows the marginal probability distribution of these relative indicesfrom one campaign across each of the sites in 2015, 2014, and 2013.The selected campaigns fall around the same time of the year for eachsite. Each line in Fig. 2 represents data from all the pixels within onecampaign for each site. Comparison of the relative contribution prob-ability density functions across the years indicates stable conditionsacross each site for the same seasons of the years. The interannualvariability is small (at least as evident in the marginal probability dis-tributions).

Fig. 3 shows the average scattering contribution of each of the threemechanisms in each site across all the campaigns. This figure showshow much of the received power in the SAR instrument is originatedfrom different scattering mechanisms and provides insight for future

retrieval algorithm development efforts on where to place the focus interms of soil moisture or vegetation parameters retrieval.

Several conclusions can be drawn from these figures:

• In dense forest (such as the LaSelva, Harvard, and Howland sites)the vegetation volume scattering is comparatively large and themode of distribution is between 55 and 65 % (Fig. 2, bottom threerows). Meanwhile, the surface contribution is relatively small (modearound 20–30 %). Thus, even though P-band has relatively highpenetration through the vegetation and surface compared to com-monly used higher frequency ranges, in dense forest much of thebackscatter is still generated by the vegetation and double-bouncescattering (in these cases 10–15 %).In the Howland and Harvard sites, which are boreal transitional/mixed forests, the mode of relative vegetation volume scattering isthe highest (65–67 %) compared to others, which results in thelowest contribution from the surface in these sites as well (Fig. 2,bottom two rows leftmost column).

• In contrast to dense forest sites, the distribution of the relative ve-getation volume scattering contribution is wider in less dense forestssuch as the Metolius, BERMS, and Duke sites. Although the mode ofthe magnitude of vegetation volume scattering is relatively high(60%), the distribution is wider compared to the dense forest sites.Meanwhile, the double-bounce scattering has a wider distributionand a higher mode which shows that the lower density of the forestallows for more interaction of the radar signal between the groundsurface and vegetation. Among the three sites Duke has a largermode for double-bounce. This can be contributed to taller treesacross this site compared to Metolius and BERMS.

• The Tonzi Ranch site has woody savanna land cover. In comparisonto Metolius, this site has an even wider distribution of vegetationvolume scattering contribution, and a smaller mode. Relative toMetolius, Tonzi shows a much smaller double-bounce scattering(higher values of double-bounce near zero) which is expected giventhat there is no vegetation cover in some parts of this site. TheChamela site (a dry subtropical forest) has a similar pattern to theTonzi site. This is likely due to the contributions of cropland andwoody savanna pixels over the Chamela flight swath (which resultsin smaller double-bounce contribution).

• Lastly, MOISST and Walnut Gulch are the two sites with the leastamount of vegetation coverage. MOISST is mostly grassland andcropland while Walnut Gulch is grassland and shrubland. As isevident in this figure, MOISST and Walnut Gulch have relativelywide distribution of surface scattering (similar to Tonzi andChamela sites) while the mode of surface scattering distribution islarger for Walnut Gulch. Both sites show a relatively wide dis-tribution of vegetation volume scattering with a mode much smallercompared to other sites.

5.2. Seasonal scattering patterns

In regions with strong seasonal variations of temperature and/orprecipitation, vegetation structure and water content can also havesignificant seasonality. This seasonality changes the interaction of radarsignal within the surface-vegetation medium (double-bounce scat-tering) as well as direct scattering from vegetation and surface. Changesin vegetation structure have a significant implication for retrieving soilmoisture or vegetation dielectric properties. Here, we focus on theTonzi Ranch site, which has strong hydroclimatic seasonality.

Tonzi Ranch has a Mediterranean climate with wet winters and drysummers. In 2014, AirMOSS had two campaigns at Tonzi Ranch, one inFebruary and one in September. These two campaigns provide the op-portunity to evaluate the impact of climate seasonality on the con-tributions from each scattering mechanism. Fig. 4 shows the decom-position results from 6 flights, 3 in each campaign season, across TonziRanch. This figure shows the absolute value of each of the Ps, Pv and Pd

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

111

to be able to compare the data across two campaigns in terms of ab-solute contributions. There is a noticeable change in the distribution ofeach of the scattering mechanisms between the two seasons.

The seasonal variations in the scattering terms can be compared toseasonal variations in soil moisture. Measurements from three in situsoil moisture probes that have been installed across the flight lines ofthe campaign and provide half-hourly measurements at seven differentdepths from 2 cm to 80 cm are used here (Romano et al., 2013; Cuencaet al., 2016). Data from 10 days prior to the first day of each campaignand during the campaign have been averaged to produce Fig. 5, whichshows the average soil moisture profile during each campaign based onin situ measurements. Soil moisture profiles have significantly changedbetween the two campaigns (as expected), with drier values occurringin September in the dry season. Moreover, the profile has more varia-bility in the wet season compared to the dry season.

Fig. 4 shows that in the September campaign all three scatteringmechanisms have decreased. The median of double-bounce, vegetationand surface scattering has decreased 16%, 25% and 36%, respectively.Two main factors contribute to the changes in the relative scatteringcontributions: vegetation structure and dielectric properties, as well assoil water content. Both factors have played a role. The soil moisturecontent is lower, and so is the direct surface backscattering. Moreover,the vegetation volume scattering is also lower, since the dry soil resultsin a drier vegetation cover and disappearance of grasslands which re-sults in less scattering surface. These two contribute to the decrease inthe double-bounce scattering which is more pronounced here. How-ever, the relative scattering contributions (shown in Fig. 6) show thatthe relative contribution of surface scattering (Ps/PT) increased by 10%in the September campaign. While the absolute scattering has de-creased, the surface scattering has more contribution to the total

Fig. 2. Probability Density Function (PDF) of relative contribution of Surface (left), Vegetation (center), and Double-bounce (right) scattering in the total observed power. Each rowrepresents data from one of the sites, and different lines represent the campaign in different years: solid line shows 2015, dashed line shows 2014, and dotted line shows 2013. Sites aresorted based on the median of the relative vegetation volume scattering, the smallest on the top row.

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

112

backscattered power. This is consistent with the fact that the decreasein vegetation cover reduces the signal attenuation and enhances signalpenetration to the ground. This comparison highlights the importanceof a dynamic vegetation volume scattering model that takes into ac-count the seasonality of vegetation structure.

5.3. Focus regions

In this section, we analyze the results from two study regions anddiscuss the spatial patterns in the decompositions. The goal is to

provide insight in the performance of the decomposition approachacross different land covers, and inherent limitations that can be im-proved in the future.

5.3.1. Case study 1: Walnut GulchThis case shows a focus region of approximately 4.6 km×8.6 km

located near Walnut Gulch in the southeastern part of the state ofArizona in US. The land cover of this region is mostly bare land andshrubs with several pivotal agricultural fields located in this domain.

Fig. 3. Average contribution of each scattering mechanism toward the total power ineach site across all campaigns. Sites are sorted based on vegetation contribution in-creasing from left to right.

Fig. 4. Decomposition results from six flights in two different campaigns (wet and dry season) at Tonzi Ranch. Columns are similar to Fig. 2, but show the absolute power contribution in[dB].

Fig. 5. Mean (solid line) and one standard deviation (shading) of soil moisture profileacross two campaign seasons at Tonzi Ranch.

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

113

We use observation from a flight campaign conducted on Aug 8, 2015.Fig. 7 shows the results of the decomposition approach in terms of therelative contribution of surface, double-bounce and volume scattering.This figure also presents an RGB-coded decomposition map whichsummarizes the contribution of each scattering mechanism in a singleimage (red: double-bounce scattering, green: volume scattering, andblue: surface scattering).

The circular patterns visible in each of the panels show differentagricultural fields that use pivotal irrigation systems. In bare land

pixels, the surface contribution is large and vegetation and double-bounce are almost zero. In vegetated pixels, however, vegetation vo-lume scattering is high, double-bounce scattering is noticeable, andsurface scattering is very small.

However, there are two pivotal systems to the right of the domain(denoted by white + signs on the bottom right panel of Fig. 7) thathave very small vegetation volume scattering and relatively largedouble-bounce scattering together with small surface scattering. Thispattern is consistent between the three flight days in a week time period

Fig. 6. Similar to Fig. 4 but showing the relative contributions.

Fig. 7. Relative contribution of surface, vegetation and double-bounce scattering mechanisms across the focus region in Walnut Gulch. Lower left panel shows the RGB-coded de-composition contributions (R: double-bounce, G: vegetation, B: surface). White pixels are either missing or filtered because of the range of incidence angle. The white crosses on the RGBpanel show two pivotal irrigation fields that are analyzed in the text. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of thisarticle.)

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

114

of this campaign (only one day is shown here). Recognizing that ve-getation volume scattering contribution is not zero, but is low com-pared to other active pivotal farms, it is possible that vegetation inthose pixels has a structure that is not consistent with the randomly-oriented model; therefore, scattering power is shifted to double-bounce.This can happen for example in vertically aligned stacks of a cereal cropthat leads to high double-bounce contribution but low random volumescattering (Jagdhuber, 2016). Analysis of the phase of the S S *HH VVacross these two fields show that the phase ranges between −14° and−25° for one of the fields, and between −32° and −55° for the secondone. The significant departure of the phase from zero shows that thepixels in these two fields are not bare soil, and support the argumentthat vegetation coverage is present and causing high double-bouncescattering, but without a random volume-like scattering contribution.

5.3.2. Case study 2: Tonzi RanchThe second focus region considered is located in Tonzi Ranch in the

central part of the state of California. It covers an area of about2.8 km×4.2 km. Observations for this region are from Feb 10, 2015.Fig. 8 shows the result of the decomposition in this domain, similar toFig. 7 for the first focus region. This domain is also dominated by bareland and low vegetated pixels that show up clearly in the decomposi-tion. However, in the left half of the domain the decomposition patternappears random, and changes dramatically between adjacent pixels.Analyzing the vegetation volume scattering contribution does not pro-vide any clue on why two adjacent regions can have such differentsurface scattering contributions.

Fig. 9 shows satellite-based imagery from the same focus regionbased on Google Earth. This image is captured within a 6 week timewindow from the time of P-band observations. Also shown is a close-upzoom of a small but representative part of the domain. Complex pat-terns of tillage are present in the left side of the domain, while there isno similar pattern on the right side. These patterns are likely con-tributing to the random pattern of surface scattering in the decom-position results. Surface roughness can also contribute to high cross-polarization observation, and this is visible in the PV/PT estimates inFig. 8. These signatures of surface roughness show that assuming con-stant surface roughness for retrievals from P-band observations evenacross small domains can result in biased estimations of surface andvegetation contribution and lead to erroneous soil moisture retrievals.

6. Discussions and conclusion

In this study, we apply a hybrid (combined model- and eigen- based)decomposition technique to P-band SAR observations. We estimate thecontributions of different scattering mechanisms across a wide range ofbiomes. The methodology builds on the previously published threecomponent decomposition approach to combine two techniques(model-based with eigen-based decomposition) and uses the coherentfully polarimetric observations. This approach is applied to polarimetricobservations from AirMOSS mission, which are phase-calibrated using adata-driven calibration scheme.

The hybrid approach does not make allometric assumptions to es-timate vegetation volume scattering. While a similar approach has beenpreviously applied to L-band observations across agricultural fields toestimate surface soil moisture (Jagdhuber et al., 2015), our study is thefirst of its kind to explore the application of a hybrid decompositionapproach with P-band observations across a wide range of vegetationcovers, in particular, dense forests. These decompositions enable futuredevelopment of retrieval algorithms to estimate soil moisture profiles inthe presence of vegetation canopy exploiting fully polarimetric ob-servations.

Applying the estimation approach to observations from 167AirMOSS campaign flights across 10 sites with diverse vegetation types,we estimated the scattering contributions of surface, vegetation, anddouble-bounce mechanisms. The relative scattering contributions ofthese three mechanisms characterize different biomes at each siteconsistent with expectations from in situ observations. Observations atWalnut Gulch showing the effects of pivotal irrigation system supportthe validity of our decomposition. The median of the vegetation volumescattering contribution ranged from 32% at Walnut Gulch to 67% atHowland, while surface scattering median ranged from 15% to 55%.The double-bounce contributions were more constant across sites, withan average median of 13% (except for Duke site with a mode of 30%).

We also investigate the results at specific site/campaigns and reportthe advantages and limitations of applying this decomposition frame-work. The difference in seasonal patterns of the estimations show thedifference in vegetation volume scattering contributions across dif-ferent seasons. Moreover, the relative contributions of surface anddouble-bounce scattering in different biomes provide valuable insightfor designing retrieval algorithms for soil moisture profile using low-frequency polarimetric observations. Such a decomposition can help to

Fig. 8. Similar to Fig. 7 but for focus region in Tonzi Ranch. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

115

better characterize the inverse problem of soil moisture profile esti-mation with fewer unknowns using one of the surface or double-bouncescattering contributions or both.

The vegetation model can be improved to include a second dipoledistribution for vegetation volume scattering and retrieving their re-lative weights. The second distribution can be targeted at tall foresttrees that have a different interaction with low-frequency microwavesignals compared to leaves and branches, and have been shown to havea higher double-bounce contribution (Moghaddam and Saatchi, 1995;Lucas et al., 2004).

Surface roughness is known to be more important in the observa-tions at low-frequencies compared to higher ones. However, estimationof surface roughness is still a challenge. As we show using a case studyat Tonzi Ranch, at P-band the effect of surface roughness may be sig-nificant and it is necessary to take into account a spatially variablesurface roughness across the domain.

Acknowledgment

Authors wish to thank members of the AirMOSS science team fortheir inputs and constructive feedbacks on different parts of this re-search. Funding for this study is provided by a subcontract toMassachusetts Institute of Technology from University of SouthernCalifornia as part of the AirMOSS mission under NASA award numberNNL12AA21C.

References

Alemohammad, S.H., Konings, A.G., Jagdhuber, T., Entekhabi, D., 2016. Characterizingvegetation and soil parameters across different biomes using polarimetric P-band SARmeasurements. In: Geoscience and Remote Sensing Symposium (IGARSS), 2016 IEEEInternational, pp. 5710–5712.

Allen, B.D., Braun, S.A., Crawford, J.H., Jensen, E.J., Miller, C.E., Moghaddam, M.,Maring, H., 2010. Proposed investigations from NASA's Earth Venture-1 (EV-1) air-borne science selections. In: Geoscience and Remote Sensing Symposium (IGARSS),2010 IEEE International, pp. 2575–2578.

Arii, M., van Zyl, J.J., Kim, Y., 2011. Adaptive model-based decomposition of polari-metric SAR covariance matrices. IEEE Trans. Geosci. Remote Sens. 49 (3),1104–1113.

Burgin, M., Clewley, D., Lucas, R.M., Moghaddam, M., 2011. A generalized radar back-scattering model based on wave theory for multilayer multispecies vegetation. IEEETrans. Geosci. Remote Sens. 49 (12), 4832–4845.

Carreiras, J.M., Quegan, S., Toan, T.L., Minh, D.H.T., Saatchi, S.S., Carvalhais, N.,Reichstein, M., Scipal, K., 2017. Coverage of high biomass forests by the ESA biomassmission under defense restrictions. Remote Sens. Environ. 196, 154–162.

Chapin, E., Chau, A., Chen, J., Heavey, B., Hensley, S., Lou, Y., Machuzak, R.,Moghaddam, M., 2012. AirMOSS: an airborne P-band SAR to measure root-zone soilmoisture. In: 2012 IEEE Radar Conference. IEEE, pp. 0693–0698.

Chapin, E., Hawkins, B.P., Harcke, L., Hensley, S., Lou, Y., Michel, T.R., Moreira, L.,Muellerschoen, R.J., Shimada, J.G., Tham, K.W., Tope, M.C., 2015. Improved abso-lute radiometric calibration of a UHF airborne radar. In: 2015 IEEE Radar Conference(RadarCon). IEEE, pp. 1720–1724.

Cloude, S., Papathanassiou, K.P., Pottier, E., 2001. Radar polarimetry and polarimetricinterferometry. IEICE Trans. Electron. 84 (12), 1814–1822.

Cloude, S., Pottier, E., 1997. An entropy based classification scheme for land applicationsof polarimetric SAR. IEEE Trans. Geosci. Remote Sens. 35 (1), 68–78.

Cloude, S.R., 2010. Polarisation: Applications in Remote Sensing. Oxford Univ. Press,Oxford, U.K.

Cloude, S.R., Pottier, E., 1996. A review of target decomposition theorems in radar

Fig. 9. Satellite-based imagery of the Tonzi Ranch study box from Google Earth. The yellow bounding box with pins at the corner shows the boundary of the study box in Fig. 8. The insetbox shows a smaller domain with finer details. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

116

polarimetry. IEEE Trans. Geosci. Remote Sens. 34 (2), 498–518.Cuenca, R.H., Hagimoto, Y., Ring, T.M., Beamer, J.P., 2016. Interpretation of in situ

observations in support of P-band radar retrievals. IEEE J. Sel. Top. Appl. Earth Obs.Remote Sens. 9 (7), 3122–3130.

Entekhabi, D., Rodriguez-Iturbe, I., Castelli, F., 1996. Mutual interaction of soil moisturestate and atmospheric processes. J. Hydrol. 184 (1-2), 3–17.

Entekhabi, D., Yueh, S., O’Neill, P., Kellogg, K., Allen, A., Bindlish, R., Brown, M., Chan,S., Colliander, A., Crow, W.T., 2014. SMAP Handbook, Number JPL Publication 400-1567. Jet Propulsion Laboratory, Pasadena, California.

Freeman, A., 2007. Fitting a two-component scattering model to polarimetric SAR datafrom forests. IEEE Trans. Geosci. Remote Sens. 45 (8), 2583–2592.

Freeman, A., Durden, S., 1998. A three-component scattering model for polarimetric SARdata. IEEE Trans. Geosci. Remote Sens. 36 (3).

Freeman, A., Durden, S.L., 1993. Three-component scattering model to describe polari-metric SAR data. In: Mott, H., Boerner, W.-M. (Eds.), Proc. SPIE 1748, RadarPolarimetry. vol. 1748. pp. 213–224.

Guissard, A., 1994. Phase calibration of polarimetric radars from slightly rough surfaces.IEEE Trans. Geosci. Remote Sens. 32 (3), 712–715.

Hajnsek, I., Jagdhuber, T., Schön, H., Papathanassiou, K.P., 2009. Potential of estimatingsoil moisture under vegetation cover by means of PolSAR. IEEE Trans. Geosci.Remote Sens. 47 (2), 442–454.

He, L., Panciera, R., Tanase, M.A., Walker, J.P., Qin, Q., 2016. Soil moisture retrieval inagricultural fields using adaptive model-based polarimetric decomposition of SARdata. IEEE Trans. Geosci. Remote Sens. 54 (8), 4445–4460.

Hensley, S., Van Zyl, J., Lavalle, M., Neumann, M., Michel, T., Muellerschoen, R., Pinto,N., Simard, M., Moghaddam, M., 2015. L-band and P-band studies of vegetation atJPL. In: 2015 IEEE Radar Conference. IEEE, pp. 516–520.

Holm, W., Barnes, R., 1988. On radar polarization mixed target state decompositiontechniques. In: Proceedings of the 1988 IEEE National Radar Conference. IEEE, pp.249–254.

Jagdhuber, T., 2012. Soil Parameter Retrieval under Vegetation Cover Using SARPolarimetry. University of Potsdam Ph.D. thesis.

Jagdhuber, T., 2016. An approach to extended fresnel scattering for modeling of depo-larizing soil-trunk double-bounce scattering. Remote Sens. 8 (10), 818.

Jagdhuber, T., Hajnsek, I., Bronstert, A., Papathanassiou, K.P., 2013. Soil moisture esti-mation under low vegetation cover using a multi-angular polarimetric decomposi-tion. IEEE Trans. Geosci. Remote Sens. 51 (4), 2201–2215.

Jagdhuber, T., Hajnsek, I., Papathanassiou, K.P., 2015. An iterative generalized hybriddecomposition for soil moisture retrieval under vegetation cover using fully polari-metric SAR. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 8 (8), 3911–3922.

Jagdhuber, T., Hajnsek, I., Sauer, S., Papathanassiou, K., Bronstert, A., 2012. SoilMoisture Retrieval Under Forest Using Polarimetric Decomposition Techniques at P-band. European Conference on Synthetic Aperture Radar (EuSAR) 2012. pp.709–712.

Kerr, Y.H., Waldteufel, P., Wigneron, J.P., Delwart, S., Cabot, F., Boutin, J., Escorihuela,M.J., Font, J., Reul, N., Gruhier, C., Juglea, S.E., Drinkwater, M.R., Hahne, A., Martin-Neira, M., Mecklenburg, S., 2010. The SMOS mission: new tool for monitoring keyelements of the global water cycle. Proc. IEEE 98 (5), 666–687.

Kim, S.-B., Moghaddam, M., Tsang, L., Burgin, M., Xu, X., Njoku, E.G., 2014. Models of L-band radar backscattering coefficients over global terrain for soil moisture retrieval.IEEE Trans. Geosci. Remote Sens. 52 (2), 1381–1396.

Kim, Y., van Zyl, J.J., 2009. A time-series approach to estimate soil moisture using po-larimetric radar data. IEEE Trans. Geosci. Remote Sens. 47 (8), 2519–2527.

Konings, A.G., Entekhabi, D., Moghaddam, M., Saatchi, S.S., 2014. The effect of variablesoil moisture profiles on P-band backscatter. IEEE Trans. Geosci. Remote Sens. 52(10), 6315–6325.

Koster, R.D., Mahanama, S.P.P., Yamada, T.J., Balsamo, G., Berg, A.A., Boisserie, M.,Dirmeyer, P.A., Doblas-Reyes, F.J., Drewitt, G., Gordon, C.T., Guo, Z., Jeong, J.-H.,Lawrence, D.M., Lee, W.-S., Li, Z., Luo, L., Malyshev, S., Merryfield, W.J.,Seneviratne, S.I., Stanelle, T., van den Hurk, B.J.J.M., Vitart, F., Wood, E.F., 2010.

Contribution of land surface initialization to subseasonal forecast skill: first resultsfrom a multi-model experiment. Geophys. Res. Lett. 37 (2).

Lee, J.-S., Pottier, E., 2009. Polarimetric Radar Imaging: From Basics to Applications. CRCPress.

Lucas, R., Moghaddam, M., Cronin, N., 2004. Microwave scattering from mixed-speciesforests, Queensland, Australia. IEEE Trans. Geosci. Remote Sens. 42 (10), 2142–2159.

McColl, K.A., Alemohammad, S.H., Akbar, R., Konings, A.G., Yueh, S., Entekhabi, D.,2017. The global distribution and dynamics of surface soil moisture. Nat. Geosci. 10(2), 100–104.

Moghaddam, M., Rahmat-Samii, Y., Rodriguez, E., Entekhabi, D., Hoffman, J., Moller, D.,Pierce, L.E., Saatchi, S., Thomson, M., 2007. Microwave Observatory of Subcanopyand Subsurface (MOSS): a mission concept for global deep soil moisture observations.IEEE Trans. Geosci. Remote Sens. 45 (8), 2630–2643.

Moghaddam, M., Saatchi, S., 1995. Analysis of scattering mechanisms in SAR imageryover boreal forest: results from BOREAS ’93. IEEE Trans. Geosci. Remote Sens. 33 (5),1290–1296.

Moghaddam, M., Saatchi, S., Cuenca, R.H., 2000. Estimating subcanopy soil moisturewith radar. J. Geophys. Res. 105 (Dll), 14899.

O’Connell, B., LaPoint, E., Turner, J., Ridley, T., Boyer, D., Wilson, A., Waddell, K., Pugh,S., Conkling, B., 2013. The Forest Inventory and Analysis Database: DatabaseDescription and User Guide ver. 5.1. 6 for Phase 2. US Department of Agriculture,Forest Service, Washington, DC.

Romano, N., D’Urso, G., Severino, G., Chirico, G., Palladino, M., Cuenca, R.H., Hagimoto,Y., Moghaddam, M., 2013. Three-and-a-half decades of progress in monitoring soilsand soil hydraulic properties. Prog. Environ. Sci. 19, 384–393.

Seneviratne, S.I., Corti, T., Davin, E.L., Hirschi, M., Jaeger, E.B., Lehner, I., Orlowsky, B.,Teuling, A.J., 2010. Investigating soil moisture-climate interactions in a changingclimate: a review. Earth Sci. Rev. 99 (3-4), 125–161.

Tabatabaeenejad, A., Burgin, M., Duan, Xueyang, Moghaddam, M., 2015. P-band radarretrieval of subsurface soil moisture profile as a second-order polynomial: firstAirMOSS results. IEEE Trans. Geosci. Remote Sens. 53 (2), 645–658.

Thompson, S.E., Harman, C.J., Konings, A.G., Sivapalan, M., Neal, A., Troch, P.A., 2011.Comparative hydrology across AmeriFlux sites: the variable roles of climate, vege-tation, and groundwater. Water Resour. Res. 47 (10).

Toan, T.L., Quegan, S., Davidson, M., Balzter, H., Paillou, P., Papathanassiou, K.,Plummer, S., Rocca, F., Saatchi, S., Shugart, H., Ulander, L., 2011. The biomassmission: mapping global forest biomass to better understand the terrestrial carboncycle. Remote Sens. Environ. 115 (11), 2850–2860 DESDynI VEG-3D Special Issue.

Truong-Loi, M.-L., Saatchi, S., Jaruwatanadilok, S., 2015. Soil moisture estimation undertropical forests using UHF radar polarimetry. IEEE Trans. Geosci. Remote Sens. 53(4), 1718–1727.

Ulaby, F.T., Long, D.G., Blackwell, W.J., Elachi, C., Fung, A.K., Ruf, C., Sarabandi, K.,Zebker, H.A., Van Zyl, J., 2014. Microwave Radar and Radiometric Remote Sensing.University of Michigan Press Ann Arbor.

van Zyl, J.J., 1989. Unsupervised classification of scattering behavior using radar po-larimetry data. IEEE Trans. Geosci. Remote Sens. 27 (1), 36–45.

van Zyl, J.J., 1992. Application of Cloude's target decomposition theorem to polarimetricimaging radar data. In: SPIE conference on Radar Polarimetry, San Diego, CA. vol.1748. pp. 184–212.

van Zyl, J.J., Arii, M., Kim, Y., 2011. Model-based decomposition of polarimetric SARcovariance matrices constrained for nonnegative eigenvalues. IEEE Trans. Geosci.Remote Sens. 49 (9), 3452–3459.

Yamaguchi, Y., Moriyama, T., Ishido, M., Yamada, H., 2005. Four-component scatteringmodel for polarimetric SAR image decomposition. IEEE Trans. Geosci. Remote Sens.43 (8), 1699–1706.

Yamaguchi, Y., Yajima, Y., Yamada, H., 2006. A four-component decomposition ofPOLSAR images based on the coherency matrix. IEEE Geosci. Remote Sens. Lett. 3(3), 292–296.

Zebker, H., Lou, Y., 1990. Phase calibration of imaging radar polarimeter Stokes matrices.IEEE Trans. Geosci. Remote Sens. 28 (2), 246–252.

S.H. Alemohammad et al. Remote Sensing of Environment 209 (2018) 107–117

117