Merging active and passive microwave observations in soil moisture data assimilation

Merging active and passive microwave observations in soil moisture data assimilation

Remote Sensing of Environment 191 (2017) 117–130 Contents lists available at ScienceDirect Remote Sensing of Environment journal homepage: www.elsev...

17MB Sizes 1 Downloads 61 Views

Remote Sensing of Environment 191 (2017) 117–130

Contents lists available at ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Merging active and passive microwave observations in soil moisture data assimilation J. Kolassa a, b, * , R.H. Reichle b , C.S. Draper b, c a

Universities Space Research Association/NPP, Columbia, MD, United States Global Modelling and Assimilation Office, NASA Goddard Spaceflight Center, Greenbelt, MD, United States c Universities Space Research Association/GESTAR, Columbia, MD, United States b

A R T I C L E

I N F O

Article history: Received 30 June 2016 Received in revised form 14 November 2016 Accepted 15 January 2017 Available online 23 January 2017 Keywords: Active/passive observation synergy Soil moisture data assimilation Microwave remote sensing

A B S T R A C T This study assesses two approaches to combine observations from active and passive satellite microwave instruments in a soil moisture data assimilation system. In the first approach, labeled ‘joint retrieval assimilation’, a single soil moisture product is simultaneously retrieved from active and passive microwave observations, and then assimilated into the NASA Catchment land surface model. In the second approach, labeled ‘separate retrieval assimilation’, separate soil moisture products are retrieved from each of the active and passive microwave observations, before being simultaneously assimilated into the model. In both approaches, a Neural Network (NN) is used to retrieve soil moisture from passive microwave AMSR-E brightness temperatures and/or active microwave ASCAT backscatter observations. A spatially distributed (3D) ensemble Kalman filter is used for the assimilation over the contiguous United States from August 2007 until September 2011. The analysis skill of both assimilation approaches is evaluated against in situ observations from 60 SCAN stations and compared to the model open loop (no assimilation) skill. When averaged across the 60 sites, the skill obtained from both assimilation experiments is very similar. For surface soil moisture, the average correlation and anomaly correlation are 0.69 and 0.58, respectively. These metrics are slightly better than those of the open loop, by 0.05 for the correlation and by 0.03 for the anomaly correlation. The root zone soil moisture estimates from the assimilation were also slightly improved compared to the open loop (by 0.03 for the average correlation and by 0.01 for the average anomaly correlation). Locally, there are differences between the skill values of the two assimilation experiments. These are related to differences in the skill of the assimilated retrieval products and how well these differences are captured by the observation errors specified in the data assimilation. © 2017 Elsevier Inc. All rights reserved.

1. Introduction Soil moisture regulates land-atmosphere coupling, and the energy and water cycles (Bateni and Entekhabi, 2012; Gentine et al., 2011; Seneviratne et al., 2006) by controlling the surface energy partitioning into sensible and latent heat fluxes. Additionally, soil moisture controls the partitioning of precipitation into runoff and infiltration (Assouline, 2013; Corradini et al., 1998; Philip, 1957) and plays a key role in the carbon cycle (McDowell, 2011; Sevanto et al., 2014). Satellite observations have proven effective at providing global soil moisture estimates with a temporal resolution of 1–2 days. In

* Corresponding author. E-mail addresses: [email protected] (J. Kolassa), [email protected] (R. Reichle), [email protected] (C. Draper).

http://dx.doi.org/10.1016/j.rse.2017.01.015 0034-4257/© 2017 Elsevier Inc. All rights reserved.

particular, microwave instruments can be used to retrieve soil moisture due their high sensitivity to the soil’s dielectric properties, which are strongly influenced by the presence of water (Dobson and Ulaby, 1986; Schmugge et al., 1986). This is reflected in the launch of the Soil Moisture and Ocean Salinity (SMOS) mission (Kerr et al., 2010) and Soil Moisture Active Passive (SMAP) mission (Entekhabi et al., 2010), which operate at L-band (1.4 GHz) and were specifically designed to observe soil moisture. Prior to these missions, soil moisture products have been retrieved from microwave sensors that were not specifically designed to measure soil moisture, including the 6.7 GHz, 10.7 GHz and 18.9 GHz channels of the passive microwave Advanced Microwave Scanning Radiometer (AMSR-E) (Owe et al., 2008, 2001; Santi et al., 2012), and the active microwave 5.3 GHz Advanced Scatterometer (ASCAT) (Notarnicola et al., 2010; Wagner et al., 2013, 1999) . In the microwave spectrum, radiometers have a higher sensitivity to soil dielectric properties and hence soil moisture than radars. However, radiometer observations are limited by a

118

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

coarse spatial resolution (∼40 km at frequencies used for soil moisture retrievals) and a strong sensitivity to other surface parameters, such as surface temperature or vegetation water content (Paloscia and Pampaloni, 1988; Schmugge et al., 1986) . Radar observations tend to be less sensitive to surface temperature effects and have a finer spatial resolution. However, their soil moisture sensitivity is less than that of radiometers and they can be subject to multiple scattering effects in topographically complex regions (Dobson and Ulaby, 1986). Their sensitivity to vegetation water content is comparable to that of radiometer observations (e.g. Dorigo et al., 2010) In addition to the single-sensor (active or passive) retrieval products, synergistic retrieval approaches have been proposed that combine the strengths of active and passive microwave observations. Most notably, the SMAP mission was designed to combine the high soil moisture sensitivity of a passive sensor with the finer spatial resolution of an active instrument flown on the same platform (Das et al., 2011, 2014; Entekhabi et al., 2010; Piles et al., 2009) . Likewise, soil moisture has also successfully been retrieved by combining observations from active and passive microwave instruments on separate platforms, including AMSR-E and ASCAT (Aires et al., 2005; Kolassa et al., 2016; Liu et al., 2012, 2011b; Santi et al., 2016). In a previous study, Kolassa et al. (2016) showed that the method chosen to combine the active and passive microwave observations in a retrieval significantly impacts the quality of the resulting soil moisture product. Their study used two approaches to combine active microwave observations from ASCAT and passive microwave observations from AMSR-E, which they labeled ‘data fusion’ and ‘a posteriori combination’. In data fusion, the active and passive microwave observations are provided as simultaneous inputs to a Neural Network (NN) retrieval algorithm and a single soil moisture product is retrieved. In the a posteriori combination, two separate retrieval products are computed with a NN retrieval algorithm, using the active and passive microwave observations as inputs, respectively. These two separate products are subsequently merged in an uncertainty-weighted averaging step. Kolassa et al. (2016) found that the data fusion method yields a better quality soil moisture product, by exploiting the complementary information in the active and passive microwave sensors. For example, a retrieval trained on a single sensor observation (e.g., on brightness temperature) will not be able to distinguish between different combinations of soil moisture and other surface variables (e.g., temperature, vegetation) that generate the same brightness temperature. Including the complementary information from another instrument (for example, backscatter observations) can help the data fusion retrieval to distinguish between these different scenarios. These results were in line with findings from similar studies (Aires et al., 2012; Kolassa et al., 2013) and reflect the main motivation for the development of multifrequency (e.g. Owe et al., 2008; Santi et al., 2012) and multi-sensor (e.g. Aires et al., 2005; Kolassa et al., 2016; Liu et al., 2011b) soil moisture retrieval products. Active and passive microwave soil moisture observations can also be combined in a data assimilation (DA) system, in this case by merging them with additional information from a land surface model. DA can generate soil moisture estimates that are more accurate than those from the observations or model alone (see e.g., Liu et al., 2011a ), while also providing soil moisture estimates between observation times, as well as estimates of the entire soil moisture profile. In the past, active/passive microwave soil moisture assimilation studies have mostly investigated the benefits of simultaneously assimilating separate active and passive retrieval products. For example, Draper et al. (2012) assimilated soil moisture products retrieved from the ASCAT and AMSR-E sensors (the same sensors as in Kolassa et al. (2016), but using physically-based retrieval algorithms rather than neural networks). They found that a simultaneous assimilation of both the active and passive soil moisture products

resulted in better model skill than the assimilation of either product alone. What Draper et al. (2012) did not investigate is whether there is a significant advantage to the simultaneous assimilation of two separate (single-sensor) retrieval products (as in their study) compared to the assimilation of a single joint active/passive retrieval. The present study builds on the work of Kolassa et al. (2016) and Draper et al. (2012) to address this question and investigate different strategies for merging active and passive sensor soil moisture information from a NN retrieval algorithm in a data assimilation system. In particular, we compare (i) the assimilation of a joint active/passive retrieval product (the data fusion product of Kolassa et al., 2016) and (ii) the assimilation of the separate active and passive retrieval products (that is, the components of the a posteriori combination method of Kolassa et al., 2016 ). The aim is to analyze whether the form in which the active and passive soil moisture data are provided to the assimilation system yields a significant difference in the skill of the assimilation output. To determine which factors cause differences between the two approaches, we will furthermore investigate how the DA system uses the satellite information in each approach. The study is organized as follows. Section 2 introduces the model and datasets used in this study and Section 3 presents the methodology employed. The results of this study are presented and discussed in Section 4. Finally, Section 5 summarizes the conclusions and discusses their implications for future studies.

2. Datasets 2.1. Model and data assimilation system The experiments use the NASA Catchment model (Koster et al., 2000) run with meteorological forcing data from the NASA ModernEra Retrospective analysis for Research and Applications (MERRA) (Rienecker et al., 2011). The precipitation forcing data were corrected using global gauge-based NOAA Climate Prediction Center Unified (CPCU) precipitation estimates and satellite and gauge based estimates from the Global Precipitation Climatology Project (GPCP) (Reichle et al., 2011; Reichle and Liu, 2014). The model was run over the contiguous United States, on the cylindrical EASE version 2 grid with a 36 km horizontal resolution (Brodzik et al., 2012) from March 2007 to September 2011. The diagnostics used to analyze the assimilation results are the daily mean surface (0–0.05 m) and root zone (0–1 m) soil moisture. The assimilation was performed using a spatially distributed (3D) Ensemble Kalman Filter. Here, 3D refers to the inclusion of nonzero horizontal correlations in the observation and model errors, which distribute the observed information to nearby model grid cells (De Lannoy and Reichle, 2016b; Reichle and Koster, 2003) . The observation operator was designed to translate the model state into surface soil moisture estimates with the same spatial support as the observations. This is approximated by the spatial convolution of the model estimates with a two-dimensional Gaussian function that contains 50% of the signal within a circle with a radius of 20 km (De Lannoy and Reichle, 2016a). The observation error variances were taken from Kolassa et al. (2016) (see Section 2.2), and the spatial correlation between the observation errors was assumed to follow a Gaussian distribution with a 0.25◦ length scale in all directions. An ensemble of 12 members was used here, since experiments indicated that increasing the ensemble size beyond 12 did not alter the conclusions of our study (not shown). The meteorological forcing and prognostic state perturbations were based on Liu et al. (2011a) and adapted for the 3D filter (see Table 1; De Lannoy and Reichle, 2016b). Lastly, to avoid spurious spatial correlations arising from the limited ensemble size, the model errors were localized by applying a distance weighting function that reduced the error correlations to

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

119

Table 1 Ensemble perturbations applied to the forcing variables – precipitation (P), downward shortwave (DSW) radiation and downward long wave (DLW) radiation – and the Catchment model prognostic variables — surface excess (srfexc) and catchment deficit (catdef). Shown are the perturbation type, which is either multiplicative (M) or additive (A), the perturbation standard deviation, the temporal and spatial correlations as well as the cross-correlations of the forcing variables. Perturbations to the prognostic variables are not cross-correlated.

P DSW DLW srfexc catdef

Type

Std dev

Temporal correlation [h]

Spatial correlation [deg]

Cross correlation with P

DSW

DLW

M M A A A

0.5 0.3 50 W m −2 0.06 kg m −2 h −1 0.15 kg m −2 h −1

24 24 24 3 3

0.2 0.2 0.2 0.2 0.2

– −0.8 0.5

−0.8 – −0.5

0.5 −0.5 –

zero beyond a radius of 0.75◦ (Gaspari and Cohn, 1999; Reichle and Koster, 2003). The ensemble was spun up for a 5 month period starting in March 2007, and the retrieval assimilation started in August 2007.

2.2. Satellite observations The daily soil moisture retrieval products assimilated in this study were derived using a statistical Neural Network (NN) approach, which was introduced in Kolassa et al. (2016) and is briefly discussed below. Specifically, three different soil moisture products were assimilated, each based on different input data: (1) AMSR-E brightness temperatures (Kawanishi et al., 2003), (2) ASCAT backscatter observations (Figa-Salda na et al., 2002), and (3) both AMSR-E brightness temperatures and ASCAT backscatter observations. The three retrieval products will be referred to as the (1) passive, (2) active, and (3) active/passive products, respectively. Where applicable, the quality control procedures described below are identical for all three retrieval products. ASCAT provides backscatter observations at C-band (5.3 GHz), whereas AMSR-E is a multi-channel instrument, of which the 6.9 GHz, 10.7 GHz, 18.7 GHz and 37 GHz brightness temperatures at horizontal and vertical polarization were used here as inputs to the NN retrieval algorithm. The typical soil penetration depth of the lower frequencies listed above is 1–2 cm, however, this can be significantly deeper for very dry soils (Prigent et al., 1999; Troch et al., 1996) . The soil moisture estimates were computed using a singlelayer feed-forward neural network trained using AMSR-E and ASCAT observations as inputs and ERA-interim/Land surface soil moisture fields (Balsamo et al., 2012, 2015) as the target data. The NN retrieval was designed not to rely on ancillary data which are either not available at the required temporal resolution (when based on visible or infrared satellite observations) or model-dependent and thus potentially problematic for data assimilation. Thus, AMSR-E and/or ASCAT observations were used as the only inputs to the NN retrieval. The neural network was trained using a classical back-propagation training algorithm (Rumelhart and Chauvin, 1995) and a Levenberg– Marquardt approach (Levenberg, 1944; Marquardt, 1963) for updating the NN weights. After training, daily mean volumetric surface soil moisture estimates were computed for the AMSR-E/ASCAT overlap period of August 2007 until September 2011 on an equal area grid with a 0.25 ◦ resolution at the equator (corresponding to a pixel size of 773 km2 ). For additional information on the NN retrieval algorithm the reader is referred to Kolassa et al. (2016). No soil moisture estimates are computed for snow covered pixels, as indicated by the Interactive Multisensor Snow and Ice Mapping System (IMS) Daily Northern Hemisphere Snow and Ice Analysis (Center, 2008), which is based on a combination of satellite observations and in situ data. The AMSR-E data have also been screened for radio frequency interference (RFI) using the spectral index technique proposed by Njoku et al. (2005). For both sensors, dense vegetation was screened by removing pixels with leaf-area-index (LAI)

greater than 3 m2 m −2 , based on an LAI climatology derived from Advanced Very High Resolution Radiometer (AVHRR) observations as used in the MERRA system (Rienecker et al., 2011). This strict threshold implies that only retrievals of relatively high quality are retained for the assimilation. Pixels with a topographic complexity in excess of 15% were removed using the index provided with the EUMETSAT ASCAT soil moisture retrieval (Wagner et al., 2013). While Draper et al. (2012) set this threshold at 10%, a 15% cutoff value was used here in order to retain more locations for which in situ measurements are available. Finally, a model based quality control was implemented in the land DA system, to remove observations during active precipitation, snow cover and/or frozen soil conditions. The three retrieval products have different skill at capturing soil moisture variations. Using the ERA-Interim/Land surface soil moisture as a reference, Kolassa et al. (2016) showed that the joint active/passive product has the best skill in terms of spatial and temporal soil moisture variations. The globally averaged error standard deviation versus ERA-Interim/Land was 0.065 m3 m −3 for the active product, 0.060 m3 m −3 for the passive product, and 0.040 m3 m −3 for the joint active/passive product (Kolassa et al., 2016). These values will be used as the observation error standard deviations in the assimilation experiments in this study (see Section 3.1). In order to assess the validity of our results when using different retrieval products, two additional soil moisture datasets are used in this study: (1) the Land Parameter Retrieval Model (LPRM) based AMSR-E soil moisture product from the Vrije Universiteit Amsterdam (Owe et al., 2008, 2001) and (2) the change detection based ASCAT surface degree of saturation observations from the Technical University Vienna (TUV) (Bartalis et al., 2008; Wagner et al., 2013, 1999) . These products were chosen because they are based on the same satellite observations as the NN retrieval products but were derived with different retrieval algorithms. The error standard deviations are 0.08 m3 m −3 for the LPRM product (as per Liu et al., 2011a and Draper et al., 2012) and ∼10% (in units of surface degree of saturation) for the TUV product (as per Draper et al., 2012 ). 2.3. In situ soil moisture observations In situ observations from the Soil Climate Analysis Network (SCAN) (Schaefer et al., 2007) are used here to evaluate the soil moisture retrieval products and assimilation output. SCAN provides observations of the volumetric soil moisture profile across the continental United States. Stations in close proximity to the coast and forested areas were removed (based on MODIS land cover type as per Draper et al., 2012 ). After an additional screening for sufficient data availability (at least 100 observations within the study period), a total of 60 stations were used for both the surface and the root zone layer. Measurements from a 5 cm depth were used to evaluate surface soil moisture, while observations from a 20 cm depth were used to evaluate root zone soil moisture. The data were quality controlled and converted to daily averages as described in De Lannoy et al. (2014).

120

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

3. Methodology 3.1. Data assimilation experiments The two data assimilation experiments compared in this study are illustrated in Fig. 1. The first experiment is the ‘joint retrieval assimilation’ of the single active/passive retrieval product. The second experiment is the ‘separate retrieval assimilation’ of both the passive product and the active products. As a baseline for the improvements obtained from the assimilation, an open loop ensemble run is performed with the same ensemble perturbations, but no assimilation. Locally, the model soil moisture and the three satellite soil moisture products have different climatologies, which necessitates a bias correction step in order to comply with the assumptions of the data assimilation system. The bias correction is implemented by matching the cumulative distribution function (CDF) of each satellite product – computed over the entire study period – to that of the Catchment model soil moisture at each grid point (CDF-matching) (Reichle and Koster, 2004). The CDF-matching is done separately for the active and passive retrieval products in the separate retrieval assimilation. For consistency with the (CDF-matched) observations being assimilated, the specified observation error standard deviation is rescaled using the ratio of standard deviations of the modeled and observed soil moisture time series at each grid cell. This approach is common practice in soil moisture data assimilation, see for example Draper et al. (2012). 3.2. Evaluation metrics The assimilation and open loop estimates are validated against SCAN in situ soil moisture observations (Section 2.3). As discussed in Section 3.1, the bias between the soil moisture retrieval products and the Catchment model is mitigated through a grid-cell level CDFmatching of the assimilated observations. The difference between

(a) Joint Retrieval Assimilation AMSR-E TB

Land Surface Model

spatial patterns in the observations and the model are removed by the CDF-matching, so that the assimilation primarily uses the temporal information in the observations. Hence, the evaluation of the retrievals and the assimilation output is based on the time series correlation and anomaly correlation metrics. The correlations (R) were computed as the Pearson correlation coefficient between the modeled and in situ soil moisture time series in each model grid cell. The anomaly correlations (Ranom ) were computed as the Pearson correlation coefficient between the anomaly time series, with the anomalies defined relative to the multi-year mean values for each calendar month over the study period. The metrics are computed at the grid cell level, yielding one value per in situ station. 4. Results In this section, we first assess the relative performance of the two assimilation experiments (Section 4.1), followed by an analysis of the assimilation skill with respect to the open loop skill (Section 4.2). 4.1. Skill of the assimilation experiments To compare the skill of the two assimilation approaches, the gray bars in Fig. 2 show the surface and root zone soil moisture metrics for both assimilation experiments averaged across the in situ sites, and the upper section of Table 2 displays the average metrics. For both the surface and root zone, the average skill of the two assimilation experiments is very similar. For the surface, the correlation for both the joint and separate retrieval assimilation is 0.69, while the average anomaly correlation for both is 0.58. For the root zone, the correlations are lower and both experiments again show a similar skill (0.67 for the correlation for both, and 0.54 and 0.52 for the anomaly correlation for the joint and separate retrieval assimilation, respectively). These results suggest that the DA system has extracted similar information from the observations in both experiments.

(b) Separate Retrieval Assimilation

ASCAT

AMSR-E TB

ASCAT

NN

NN

NN

SM active/ passive

SM passive

SM active

Data assimilation

SM analysis

Land Surface Model

Data assimilation

SM analysis

Fig. 1. Schematic illustrating (a) joint retrieval assimilation and (b) separate retrieval assimilation. TB refers to brightness temperature observations and s refers to backscatter observations.

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

(b)

surface 0.8

0.8

0.7

0.7

0.6

0.6

correlation [-]

correlation [-]

(a)

0.5

0.4

0.3

0.3

correlation

anomaly correlation

root zone

0.5

0.4

0.2

121

0.2

correlation

anomaly correlation

Fig. 2. Correlation and anomaly correlation for the assimilation and open loop experiments, averaged over all sites, for (a) surface and (b) root-zone soil moisture. Error bars indicate 95% confidence intervals. The metrics are computed using the daily average soil moisture fields produced on the EASE 36km grid using the DA system.

Table 2 Spatially averaged temporal correlation (R) and anomaly correlation (Ranom ) for surface and root zone soil moisture estimates from the open loop and assimilation experiments. Skill metrics are computed against SCAN in situ measurements and using the daily average soil moisture fields produced on the EASE 36 km grid using the DA system. Experiment

Joint retrieval assimilation Separate retrieval assimilation Open loop LPRM & TU Vienna assimilation

Surface

between the two assimilation runs and illustrates the spatial variations in the relative skill. Absolute difference values are mostly below 0.03, but occasionally reach up to 0.05. In general, the differences are not consistent across all metrics and/or soil layers, with the exception of some regions in which one of the assimilation approaches is consistently superior. The first such region is a narrow strip in the western US, stretching from eastern Montana through Idaho, Utah and Arizona, labeled region A in Fig. 4 (a). In this region, the separate retrieval assimilation shows a better skill than the joint retrieval assimilation in the surface layer. These local differences are related to a combination of three factors: (i) the skill of the assimilated retrievals, (ii) the relative weights given to the observations, and (iii) the number of assimilated observations. Fig. 5 addresses the first of the three factors by showing maps of the correlation and anomaly correlation skill metrics of the joint retrieval product versus SCAN soil moisture observations, together with maps of the skill difference between each of the two singlesensor retrieval products and the joint retrieval product skill. In region A the joint and passive retrieval products have a similar skill on average, with skill differences that rarely exceed the correlation significance level (Fig. 3) and that are not consistent across the region

surface 0.8 0.7

correlation [-]

A key driver of the assimilation performance is the skill of the assimilated observations. Fig. 3 and Table 3 show the correlation and anomaly correlation of the retrieved soil moisture with respect to the in situ observations. The joint retrieval has a correlation of 0.46, which is slightly higher than that of the passive (0.44) or active (0.41) products. The anomaly correlations of the joint and active retrieval (0.45 and 0.48, respectively) are higher than for the passive retrieval product (0.37). For both metrics, the average skill across the active and passive products is less than the skill of the joint retrieval product. This is consistent with Kolassa et al. (2016), who showed that the optimal average of the two single-sensor retrieval products had a lower skill than a single joint retrieval. Table 2 shows that the skill of the assimilation estimates from both experiments is virtually the same, regardless of whether the DA system ingested the joint retrieval product or assimilated the two single-sensor products simultaneously. This result is different from what Kolassa et al. (2016) found for the merging of active and passive information at the retrieval level. There, the joint approach performed better, because it used the complete and complementary information in the active and passive microwave observations (that is, backscatter and brightness temperature). The assimilation system, by contrast, can better identify which of the two separate retrievals is more correct when they disagree, by using additional and (mostly) independent information from the model. Thus, even with an imperfect knowledge of the observation and model error estimates, the assimilation of one joint active/passive retrieval and the simultaneous assimilations of two single-sensor retrievals yield equally skillful soil moisture estimates. While the average skill metrics for the two assimilation experiments are very similar, there are some local skill differences. Fig. 4 shows maps of the correlation and anomaly correlation difference

0.6

0.5

0.4

0.3

Root zone

R

Ranom

R

Ranom

0.69 0.69 0.64 0.72

0.58 0.58 0.55 0.62

0.67 0.67 0.64 0.68

0.54 0.52 0.52 0.54

0.2

correlation

anomaly correlation

Fig. 3. Correlation and anomaly correlation for the three surface soil moisture retrieval products averaged over all sites. Error bars indicate 95% confidence intervals. The metrics are computed using the satellite soil moisture products with a temporal resolution of 2–3days and provided on a 25km equal area grid.

122

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130 Table 3 Same as Table 2, but for surface soil moisture retrieval products with a temporal resolution of 2–3 days and provided on a 25 km equal area grid. Retrieval

R

Ranom

Active/passive Passive Active LPRM (X-band) TU Vienna

0.46 0.44 0.41 0.51 0.50

0.45 0.37 0.48 0.47 0.52

(Fig. 5c, d). The active retrieval product has correlations that are lower by ∼0.2 for most stations in region A (Fig. 5e) and anomaly correlations that are slightly smaller than those of the joint retrieval product (Fig. 5f). Before relating the results of Fig. 5 to the local skill differences seen in Fig. 4, we address the second factor, the relative impact of the assimilated observations and the model forecast in the analysis. This impact can be measured by the absolute ratio of the observation-minus-analysis (O − A) residuals (or differences) to the observation-minus-forecast (O − F) residuals. In the scalar case, this metric is equal to (1 − K), where K is the DA system’s Kalman gain, which modulates the translation of the O − F residuals (or innovations) into the analysis increments   (O − A) applied to the forecast soil   moisture. Lower values of  O−A O−F  indicate that the analysis soil moisture is closer to the observations than to the forecast soil moisture and thus signal a higher of the observations in the analysis.  weight    The time mean of the  O−A O−F  ratio is shown in Fig. 6.

For most stations in region A Fig. 6 shows that the active  retrieval   is given less weight in the assimilation (higher  O−A O−F ) than the joint or passive retrieval product. This is a direct consequence of the higher error standard deviation of the active retrieval product. Finally, Fig. 7 addresses the third factor by showing the average number of data points assimilated per day at each location. The figure shows that in region A fewer active retrieval observations are assimilated compared to the other two retrieval products. Comparing Figs. 5 and 6, the better performance of the separate retrieval assimilation in region A likely results from the observation error specification correctly reflecting the higher passive retrieval skill in this region, thus giving the passive observations more weight. In the joint retrieval assimilation mode, the retrieval product has a single observation error that does not distinguish between the two sensors and thus performs less well at isolating the reliable retrieval information. For the Eastern US (region B) Fig. 4 shows that the joint retrieval assimilation has a better correlation and a slightly better anomaly correlation than the separate retrieval assimilation in both layers. A comparison with the observation skill in Fig. 5 shows that in this region the active retrieval has a correlation comparable to that of the joint active/passive retrieval, while also having a higher anomaly correlation. The passive retrieval has a similar correlation, but a lower anomaly correlation compared to the joint active/passive retrieval. However, Fig. 6 shows that in this region the active retrieval product     is given a similar or lower observation weight (higher  O−A O−F ) compared the passive product, again due to the higher observation error specified for the active data. The separate retrieval assimilation then gives more weight to the less reliable passive retrieval information

(a)

(b)

(c)

(d)

Fig. 4. Skill difference between assimilation experiments, computed as joint retrieval assimilation skill metric minus separate retrieval assimilation skill metric, for (a, c) correlations and (b, d) anomaly correlations for (a, b) surface and (c, d) root zone soil moisture. Red colors indicate a higher skill of the joint retrieval assimilation and blue colors indicate a higher skill of the separate retrieval assimilation. The regions discussed in Section 4.1 are shown in panel (a).

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

(a)

(b)

(c)

(d)

(c)

(d)

123

Fig. 5. Skill of the retrieval products with respect to the in situ surface soil moisture. (a) Correlation and (b) anomaly correlation of the joint active/passive retrieval product. (c, e) Correlation differences and (d, f) anomaly correlation differences with respect to the joint active/passive retrieval for (c, d) the passive retrieval product and (e, f) the active retrieval product. In the difference maps, red colors indicate a higher skill of the joint retrieval product and blue colors indicate a higher skill of the individual retrieval product.

and less weight to the more reliable active retrieval information, in this region. By contrast, the joint retrieval has a good skill and a higher impact than both single-sensor retrieval products, resulting in the better performance of the joint retrieval assimilation in the Eastern US. In summary, both assimilation approaches have a similar skill averaged over CONUS, suggesting that the DA system extracts similar information from the retrievals, regardless of whether the information is provided in the form of a joint active/passive retrieval product or as two single-sensor retrieval products. Locally, the choice of assimilation approach can substantially impact the analysis skill, with the relative performance of the assimilation approaches depending on the observation error specifications and how well these reflect the actual retrieval skill. In regions where the specified observation errors accurately capture the relative skill of the assimilated observations, the separate retrieval assimilation is better

able to isolate the reliable retrieval information, yielding a higher analysis skill. If more sophisticated and localized observation errors estimates were available (as opposed to the spatially constant estimates used here (Section 2.2)), the separate retrieval assimilation method should then outperform the joint assimilation, by better separating reliable and unreliable retrieval information in the observations. While the model error estimates are equally relevant to the analysis skill, the same ensemble perturbations were used in each experiment here and so are not expected to contribute to differences in the performance of each assimilation experiment. 4.2. Assimilation and open loop skill Next, we investigate the skill of the two assimilation experiments with respect to the model open loop (no data assimilated). Besides the assimilation skill metrics, Fig. 2 also shows the surface

124

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

(a)

(b)

(c)

    Fig. 6. Mean  O−A O−F  for the (a) active/passive retrieval data during joint retrieval assimilation, (b) passive retrieval data during separate retrieval assimilation, and (c) active retrieval data during separate retrieval assimilation. Low numbers (that  is, warm colors) indicate larger impact of observations, high numbers (that is, cold colors) indicate a larger   impact of the forecast. The plot titles indicate the spatial average  O−A O−F  value for each retrieval product in parentheses.

soil moisture correlation and anomaly correlation for the open loop, with values of 0.64 and 0.55 respectively (summarized in Table 2). For the root zone, the corresponding open loop values are 0.64 and 0.52. Fig. 2 shows that, on average, the assimilation experiments have a slightly higher skill than the open loop. The correlation

and anomaly correlation improvements for both assimilation experiments are 0.05 and 0.03, respectively, in the surface layer, and 0.03 and 0–0.02 in the root zone (Table 2). This net skill improvement is also shown in a scatter plot of the assimilation and open loop skill values in Fig. 8. Furthermore, Fig. 9 illustrates the spatial patterns

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

125

(a)

(b)

(c)

Fig. 7. Mean number of data points assimilated per day, for the assimilation of (a) the joint active/passive retrieval product, (b) the passive retrieval product, and (c) the active retrieval product.

of the correlation and anomaly correlation difference between each assimilation experiment and the open loop for surface soil moisture. While these differences are small, the assimilation improves the model at most locations. Stations with a skill degradation at the surface are close to mountain ranges, where microwave-based soil moisture retrievals are known to be uncertain, and where the in situ

data can also be less representative. Generally, the largest improvements in the surface are obtained in the Eastern part of the US and the Central Midwest, which are characterized by higher latent heat flux and soil moisture compared to the Western US. For the root zone, the skill improvements are smaller (Fig. 10). A degradation of the skill occurs for some stations, which are primarily

126

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

(a)

(b)

surface

1

joint retrieval DA joint retrieval separate retrieval DA

0.9

0.9

0.8

0.8

open loop skill (Ranom)

open loop skill (Ranom)

1

0.7 0.6 0.5 0.4

0.6 0.5 0.4 0.3

0.2

0.2

0.1

0.1 0.2

0.4

0.6

0.8

1

joint DA jointretrieval retrieval separate retrieval DA

0.7

0.3

0 0

root zone

0 0

assimilation skill (Ranom)

0.2

0.4

0.6

0.8

1

assimilation skill (Ranom)

Fig. 8. (a) Surface and (b) root zone soil moisture skill of the assimilation estimates versus that of the open loop. Skill is defined as the anomaly correlation obtained with respect to the in situ data for the joint retrieval assimilation experiment (blue squares) and the separate retrieval assimilation experiment (red circles).

located in the Western US. For stations in the topographically complex terrain near the Rocky Mountains, this is related to the higher uncertainty of the microwave retrievals. The topographic complexity

for the stations ranges between 10% and 15% and thus the degradation could be mitigated through a stricter topography flag. For the remainder of the stations, where the assimilations improves the

(a)

(b)

(c)

(d)

Fig. 9. Surface soil moisture skill improvement for (a, b) joint retrieval assimilation and (c, d) separate retrieval assimilation, measured in terms of (a, c) R and (b, d) Ranom . Improvement is computed as the assimilation skill metric minus the open loop skill metric. Red colors indicate an improvement of the assimilation over the open loop skill and blue colors indicate a degradation.

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

(a)

(b)

(c)

(d)

127

Fig. 10. Same as Fig. 9, but for root zone soil moisture.

surface soil moisture, but degrades the root zone soil moisture, this is likely related to a misrepresentation of the physical processes in the model. The reduced skill and/or degradation for some stations in the root zone has been observed in previous soil moisture assimilation studies (De Lannoy and Reichle, 2016b; Reichle et al., 2016). The skill improvements from the assimilation of the retrieval products are lower than those found by Draper et al. (2012), but are nevertheless generally positive across the domain for the surface soil moisture. The small improvements are related to the relative skill of the retrieval products compared to the model, which is shown in Fig. 11. Fig. 11 also indicates where the anomaly correlation difference, Ranom,model − Ranom,retrieval , is 0 (solid line) and 0.2 (dashed line), respectively. For a large number of stations the model has an anomaly correlation that is higher by 0.2 or more than that of the retrieval, which approximates the maximum anomaly correlation difference identified by Reichle et al. (2008) and Draper et al. (2012) for the assimilation to be beneficial. The retrieval skill is lower relative to the model skill in this study than it was in Draper et al. (2012) (compare to their Fig. 4). In part, this is due to improvements in the model skill, open loop surface and root zone anomaly correlations are 0.55 and 0.52 here, compared to 0.47 and 0.45 in Draper et al. (2012). This is associated with recent model updates (De Lannoy et al., 2014), and the use of observation-corrected precipitation (Reichle and Liu, 2014). To determine the impact of the NN retrieval skill, we compared the NN retrieval to the LPRM and TU Vienna soil moisture retrieval products used in Draper et al. (2012). The passive microwave LPRM retrieval correlation and anomaly correlation against the surface the in situ data are 0.51 and 0.47, respectively, compared to 0.44 and 0.37, respectively, for the passive NN retrieval product (Table 3). The active microwave TU Vienna retrieval product has a correlation and anomaly correlation of 0.50 and 0.52, respectively, compared to 0.41 and 0.48, respectively, for the active NN retrieval product (Table 3).

The lower skill of the NN retrievals could be related to the fact that the NN retrievals rely purely on the microwave sensor information and do not use ancillary data products to account for surface temperature or vegetation cover effects. Residual RFI contamination of the AMSR-E 6.9 GHz brightness temperatures, which are used in the NN retrieval but not in the LPRM product, could also contribute to the lower skill. Draper et al. (2012) thus assimilated generally more skillful retrievals into a generally less skillful model, hence their skill improvement from the assimilation was greater. As an additional test, we also simultaneously assimilated the LPRM and TU Vienna retrievals into our current assimilation system. The resulting surface soil moisture correlation and anomaly correlation values are 0.72 and 0.62 (compared to 0.69 and 0.58 for the NN retrieval assimilation), respectively, and the corresponding root zone metrics are 0.68 and 0.54 (compared to 0.67 and 0.52 for the NN retrieval assimilation), respectively (Table 2). Despite their higher skill, the LPRM and TU Vienna retrievals are not suited to compare the joint and separate retrieval assimilation methods. While the ESA-CCI product (Liu et al. 2011b,Liu et al., 2012), can be considered a joint retrieval that merges the LPRM and TUV data, over most of CONUS the ESA-CCI algorithm primarily uses only one of its (active or passive) input products, rather than generating a truly joint active/passive estimate (cf. Fig. 10 in Liu et al., 2011b). Thus, despite their generally lower skill, the NN retrievals are better suited for the main purpose of our study, which is to compare the relative skill of the joint and separate retrieval assimilation approaches. 5. Conclusions and perspectives This study investigated the extent to which the method chosen to combine data from active and passive microwave instruments in

128

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

1 0.9

active/passive retrieval passive retrieval active retrieval

retrieval skill (Ranom)

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0

0.2

0.4

0.6

0.8

1

model skill (Ranom) Fig. 11. Model (open loop) skill versus retrieval skill for the anomaly correlations obtained with respect to the in situ data for the active/passive retrieval product (red squares), the passive retrieval product (blue diamonds) and the active retrieval product (green circles). Also shown are the 1:1 correspondence line (solid), as well as the line separating points with a retrieval skill that is lower by 0.2 or more compared to the model skill (dashed).

a soil moisture assimilation system impacts the model skill gained from those observations. The assimilation of a joint active/passive soil moisture retrieval, and the simultaneous assimilation of separate active and passive microwave soil moisture retrievals were investigated. Averaged over CONUS, the results show little difference in the analysis skill from each approach. Both methods are able to use the information provided by the active and passive instrument, although they do this at different stages. The joint retrieval assimilation method exploits information complementarity at the retrieval level, yielding a higher quality soil moisture product to be assimilated. The separate retrieval assimilation technique uses the additional information from the model to modulate the impact of the two separate retrievals. The results indicate that, despite imperfect knowledge of the observation errors, a data assimilation system can extract the complementary information provided by both satellite sensors, regardless of the form in which that information is provided. While both approaches yield a similar domain-averaged performance, there were differences at the local scale, which highlight differences between the assimilation approaches. Generally, the skill of the two assimilation approaches depends on the retrieval skill, the specified observation error standard deviation and the number of observations assimilated. The separate retrieval assimilation performed better in regions where the choice of the error standard deviation for each retrieval product correctly reflected their relative skill. In contrast, in regions where the specified observation errors did not accurately reflect the relative skill of the two separate retrieval products, the joint retrieval assimilation performed better due to the generally higher skill of the active/passive retrieval product. Therefore, the use of observation error estimates that accurately reflect the relative errors of the two separate retrievals locally, would likely result in a better overall performance of the separate retrieval

assimilation, which offers more flexibility in treating reliable and unreliable information from each sensor independently. Both assimilation approaches yield small, but generally positive, improvements in the analysis skill compared to the open loop. On average, the surface soil moisture correlation increases by 0.05 and the surface anomaly correlation increases by 0.03, whereas the root zone correlation increases by 0.03, and the root zone anomaly correlation increases by 0.02. The assimilation is found to improve the model surface skill across most in situ stations used in the evaluation. The main exceptions are stations in the proximity of mountainous regions, where retrievals based on microwave observations are known to have larger errors. In the root zone, skill improvements occur in the Eastern US, whereas stations in the Western US showed a skill degradation. In this study, the observation error was specified to be spatially uniform for each retrieval product, and was subsequently scaled within the observation bias correction using the ratio of the local modeled and observed soil moisture variability (standard deviation). This approach can produce unrealistic spatial patterns in the observation error, and can have the unintended effect of giving more weight to the retrievals where they are very noisy. This includes, for example, regions in the proximity of mountain ranges, in which microwave retrieval products are known to be less reliable. A more refined approach to specifying the observation errors (e.g. Draper et al., 2013) and thus better separating reliable from unreliable information should help to improve the analysis skill obtained from satellite soil moisture assimilation. Additionally, the pixel-level CDF-matching that is applied as a bias correction step in this study removes the spatial information in the retrieval products, potentially removing useful information on spatial variability. A future study will investigate the use of a statistical retrieval approach trained on the same model as used in the assimilation. This approach would use global model soil moisture to train the NN retrieval algorithm, thereby producing soil moisture retrieval estimates within the model climatology. This approach may reduce the need for a localized bias correction (CDF-matching) and potentially enhance the amount of soil moisture information extracted during data assimilation. Acknowledgments J. Kolassa was supported by an appointment to the NASA Postdoctoral Program at the Goddard Spaceflight Center, administered by Universities Space Research Association under contract with NASA. Additional funding was provided by the NASA Soil Moisture Active Passive mission. Computational resources for this study were provided by the NASA High-End Computing (HEC) Program through the NASA Center for Climate Simulation (NCCS) at the Goddard Space Flight Center. The authors would like to thank the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT) for providing ASCAT backscatter observations, the National Snow and Ice Data Center (NSIDC) for providing AMSR-E brightness temperature data and the National Water and Climate Center for providing SCAN in situ soil moisture observations. Additional thanks are extended to Gabrielle De Lannoy and Qing Liu for preprocessing and quality control of the in situ observations. The ERA-interim/Land soil moisture fields used in the retrieval algorithm calibration are provided publicly by the European Centre for Medium-Range Weather forecasts for the period 2007–2010 and additional data for the year 2011 was provided by Clement Albergel. The ASCAT soil moisture data were provided by the Department of Geodesy and Geoinformation at the Technical University of Vienna using the WARP 5.4 retrieval algorithm with a spatial resolution of 12.5 km. The LPRM soil moisture data were provided by the Faculty of Earth and Life Science Department of Eco-Hydrology at the Vrije Universiteit Amsterdam.

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

References Aires, F., Aznay, O., Prigent, C., Paul, M., Bernardo, F., 2012. Synergistic multiwavelength remote sensing versus separate retrieval assimilation of retrieved products: application for the retrieval of atmospheric profiles using MetOp-A. J. Geophys. Res 117 (D18), 2012. http://dx.doi.org/10.1029/2011JD017188. Aires, F., Prigent, C., Rossow, W.B., 2005. Sensitivity of satellite microwave and infrared observations to soil moisture at a global scale: 2. Global statistical relationships. J. Geophys. Res. 110, D11103. http://dx.doi.org/10.1029/2004JD005094. Assouline, S., 2013. Infiltration into soils: conceptual approaches and solutions. Water Resour. Res. 49 (4), 1755–1772. http://dx.doi.org/10.1002/wrcr.20155. Balsamo, G., Albergel, C., Beljaars, A., Boussetta, S., Brun, E., Cloke, H., et al. 2012. ERA-Interim/Land: A Global Land-Surface Reanalysis Based on ERA-Interim Meteorological Forcing. ERA Report Series, ECMWF. Balsamo, G., Albergel, C., Beljaars, A., Boussetta, S., Brun, E., Cloke, H., Dee, D., Dutra, E., Mu noz Sabater, J., Pappenberger, F., De Rosnay, P., 2015. ERA-interim/land: a global land surface reanalysis data set. Hydrol. Earth Syst. Sci. 19 (1), 389–407. http://dx.doi.org/10.5194/hess-19-389-2015. Bartalis, Z., Naeimi, V., Hasenauer, S., Wagner, W., 2008. ASCAT Soil Moisture Product Handbook. ASCAT Soil Moisture Report Series., pp. 15. Bateni, S.M., Entekhabi, D., 2012. Relative efficiency of land surface energy balance components. Water Resour. Res. 48 (4). http://dx.doi.org/10.1029/ 2011WR011357. Brodzik, M.J., Billingsley, B., Haran, T., Raup, B., Savoie, M.H., 2012. EASE-grid 2.0: incremental but significant improvements for Earth-gridded data sets. ISPRS Int. J. Geo-Inf. 1 (1), 32–45. http://dx.doi.org/10.3390/ijgi1010032. Center, N.I., 2008. Updated Daily. IMS Daily Northern Hemisphere Snow and Ice Analysis at 1 Km, 4 Km, and 24 Km Resolutions. National Snow and Ice Data Center. Digital media, Boulder, CO. Corradini, C., Morbidelli, R., Melone, F., 1998. On the interaction between infiltration and Hortonian runoff. J. Hydrol. 204 (1), 52–67. http://dx.doi.org/10.1016/S00221694(97)00100-5. Das, N.N., Entekhabi, D., Njoku, E.G., 2011. An algorithm for merging SMAP radiometer and radar data for high-resolution soil-moisture retrieval. IEEE Trans. Geosci. Remote Sens. 49 (5), 1504–1512. http://dx.doi.org/10.1109/TGRS.2010.2089526. Das, N.N., Entekhabi, D., Njoku, E.G., Shi, J.J., Johnson, J.T., Colliander, A., 2014. Tests of the SMAP combined radar and radiometer algorithm using airborne field campaign observations and simulated data. IEEE Trans. Geosci. Remote Sens. 52 (4), 2018–2028. De Lannoy, G.J., Koster, R.D., Reichle, R.H., Mahanama, S.P., Liu, Q., 2014. An updated treatment of soil texture and associated hydraulic properties in a global land modeling system. J. Adv. Model. Earth Syst. 6 (4), 957–979. http://dx.doi.org/10. 1002/2014MS000330. De Lannoy, G.J., Reichle, R., 2016a. Assimilation of SMOS brightness temperatures or soil moisture retrievals into a land surface model. Hydrol. Earth Syst. Sci. Discuss http://dx.doi.org/10.5194/hess-2016-414. De Lannoy, G.J., Reichle, R.H., 2016b. Global assimilation of multi-angle and multipolarization SMOS brightness temperature observations into the GEOS-5 catchment land surface model for soil moisture estimation. J. Hydrometeorol. http:// dx.doi.org/10.1175/JHM-D-15-0037.1. Dobson, M.C., Ulaby, F.T., 1986. Active microwave soil moisture research. IEEE Trans. Geosci. Remote Sens. (1), 23–36. http://dx.doi.org/10.1109/TGRS.1986.289585. Dorigo, W., Scipal, K., Parinussa, R., Liu, Y., Wagner, W., De Jeu, R., Naeimi, V., 2010. Error characterisation of global active and passive microwave soil moisture datasets. Hydrol. Earth Syst. Sci. 14 (12), 2605–2616. http://dx.doi.org/10.5194/ hess-14-2605-2010. Draper, C., Reichle, R., de Jeu, R., Naeimi, V., Parinussa, R., Wagner, W., 2013. Estimating root mean square errors in remotely sensed soil moisture over continental scale domains. Remote Sens. Environ. 137, 288–298. http://dx.doi.org/10.1016/j.rse. 2013.06.013. Draper, C.S., Reichle, R.H., De Lannoy, G.J.M., Liu, Q., 2012. Assimilation of passive and active microwave soil moisture retrievals. Geophys. Res. Lett. 39 (4). http://dx. doi.org/10.1029/2011GL050655. Entekhabi, D., Njoku, E.G., O’Neill, P.E., Kellogg, K.H., Crow, W.T., Edelstein, W.N., 2010. The soil moisture active passive (SMAP) mission. Proc. IEEE 98 (5), 704–716. http://dx.doi.org/10.1109/JPROC.2010.2043918. Figa-Salda na, J., Wilson, J.J., Attema, E., Gelsthorpe, R., Drinkwater, M.R., Stoffelen, A., 2002. The advanced scatterometer (ASCAT) on the meteorological operational (MetOp) platform: a follow on for European wind scatterometers. Can. J. Remote. Sens. 28 (3), 404–412. http://dx.doi.org/10.5589/m02-035. Gaspari, G., Cohn, S.E., 1999. Construction of correlation functions in two and three dimensions. Q. J. R. Meteorol. Soc. 125 (554), 723–757. http://dx.doi.org/10.1002/ qj.49712555417. Gentine, P., Polcher, J., Entekhabi, D., 2011. Harmonic propagation of variability in surface energy balance within a coupled soil–vegetation–atmosphere system. Water Resour. Res. 47 (5). http://dx.doi.org/10.1029/2010WR009268. Kawanishi, T., Sezai, T., Ito, Y., Imaoka, K., Takeshima, T., Ishido, Y., Shibata, A., Miura, M., Inahata, H., Spencer, R., 2003. The advanced microwave scanning radiometer for the earth observing system (AMSR-e), NASDA’s contribution to the EOS for global energy and water cycle studies. IEEE Trans. Geosci. Rem. Sens. 41, 184–194. http://dx.doi.org/10.1109/TGRS.2002.808331. 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., MartinNeira, M., Mecklenburg, S., 2010. The SMOS mission: new tool for monitoring key elements of the global water cycle. Proc. IEEE 98 (5), 666–687. http://dx.doi.org/ 10.1109/JPROC.2010.2043032.

129

Kolassa, J., Aires, F., Polcher, J., Prigent, C., Jimenez, C., Pereira, J.M., 2013. Soil moisture retrieval from multi-instrument observations: information content analysis and retrieval methodology. J. Geophys. Res 118 (10), 4847–4859. http://dx.doi.org/10. 1029/2012JD018150. Kolassa, J., Gentine, P., Prigent, C., Aires, F., 2016. Soil moisture retrieval from AMSR-e and ASCAT microwave observation synergy. Part 1: satellite data analysis. Remote Sens. Environ. 173, 1–14. http://dx.doi.org/10.1016/j.rse.2015.11.011. Koster, R.D., Suarez, M.J., Ducharne, A., Stieglitz, M., Kumar, P., 2000. A Catchment-Based Approach to Modeling Land Surface Processes in a General Circulation Model 1. Model Structure. Levenberg, K., 1944. A Method for the Solution of Certain Non-Linear Problems in Least Squares. , pp. 164–168. Liu, Q., Reichle, R.H., Bindlish, R., Cosh, M.H., Crow, W.T., de Jeu, R., De Lannoy, G.J., Huffman, G.J., Jackson, T.J., 2011a. The contributions of precipitation and soil moisture observations to the skill of soil moisture estimates in a land data assimilation system. J. Hydrometeorol. 12 (5), 750–765. http://dx.doi.org/10.1175/JHMD-10-05000.1. Liu, Y.Y., Dorigo, W.A., Parinussa, R.M., de Jeu, R.A.M., Wagner, W., McCabe, M.F., Evans, J.P., van Dijk, A.I.J.M., 2012. Trend-preserving blending of passive and active microwave soil moisture retrievals. Remote Sens. Environ. 123, 280–297. http:// dx.doi.org/10.1016/j.rse.2012.03.014. Liu, Y.Y., Parinussa, R.M., Dorigo, W.A., De Jeu, R.A.M., Wagner, W., Van Dijk, A.I.J.M., Evans, J.P., 2011b. Developing an improved soil moisture dataset by blending passive and active microwave satellite-based retrievals. Hydrol. Earth Syst. Sci. 15 (2), 425–436. http://dx.doi.org/10.5194/hess-15-425-2011. Marquardt, D.W., 1963. An algorithm for least-squares estimation of nonlinear parameters. J. Soc. Ind. Appl. Math. 11 (2), 431–441. http://dx.doi.org/10.1137/0111030. McDowell, N.G., 2011. Mechanisms linking drought, hydraulics, carbon metabolism, and vegetation mortality. Plant Physiol. 155 (3), 1051–1059. http://dx.doi.org/10. 1104/pp.110.170704. Njoku, E.G., Ashcroft, P., Chan, T.K., Li, L., 2005. Global survey and statistics of radiofrequency interference in AMSR-e land observations. IEEE Trans. Geosci. Remote Sens. 43 (5), 938–947. http://dx.doi.org/10.1109/TGRS.2004.837507. Notarnicola, C., Santi, E., Brogioni, M., Paloscia, S., Pettinato, S., Preziosa, G., Ventura, B., 2010. Neural network adaptive algorithm applied to high resolution c-band SAR images for soil moisture retrieval in bare and vegetated areas. In Remote Sensing. International Society for Optics and Photonics., pp. 78290F.. October. Owe, M., de Jeu, R., Holmes, T., 2008. Multi-sensor historical climatology of satellitederived global land surface moisture. J. Geophys. Res.-Earth. 113. http://dx.doi. org/10.1029/2007JF000769. Owe, M., deJeu, R., Walker, J., 2001. A methodology for surface soil moisture and vegetation optical depth retrieval using the microwave polarization difference index. IEEE Trans. Geosci. Remote Sens. 39, 1643–1654. http://dx.doi.org/10.1109/36. 942542. Paloscia, S., Pampaloni, P., 1988. Microwave polarization index for monitoring vegetation growth. IEEE Trans. Geosci. Remote Sens. 26 (5), 617–621. Philip, J.R., 1957. The theory of infiltration: 5. The influence of the initial moisture content. Soil Sci. 84 (4), 329–340. Piles, M., Entekhabi, D., Camps, A., 2009. A change detection algorithm for retrieving high-resolution soil moisture from SMAP radar and radiometer observations. IEEE Trans. Geosci. Remote Sens. 47 (12), 4125–4131. http://dx.doi.org/10.1109/TGRS. 2009.2022088. Prigent, C., Rossow, W.B., Matthews, E., Marticorena, B., 1999. Microwave radiometric signatures of different surface types in deserts. J. Geophys. Res. Atmos. (19842012) 104 (D10), 12147–12158. http://dx.doi.org/10.1029/1999JD900153. Reichle, R.H., Crow, W.T., Koster, R.D., Sharif, H.O., Mahanama, S.P.P., 2008. Contribution of soil moisture retrievals to land data assimilation products. Geophys. Res. Lett. 35 (1). http://dx.doi.org/10.1029/2007GL031986. Reichle, R.H., De Lannoy, G.J.M., Liu, Q., Ardizzone, J.V., Chen, F., Colliander, A., Conaty, A., Crow, W., Jackson, T., Kimball, J., Koster, R.D., Brent Smith, E., 2016. Soil moisture active passive mission l4 SM data product assessment (version 2 validated release). GMAO Office Note No. 12 (Version 1.0), 55. Reichle, R.H., Koster, R.D., 2003. Assessing the impact of horizontal error correlations in background fields on soil moisture estimation. J. Hydrometeorol. 4 (6), 1229– 1242. http://dx.doi.org/10.1175/1525-7541(2003)004. Reichle, R.H., Koster, R.D., 2004. Bias reduction in short records of satellite soil moisture. Geophys. Res. Lett. 31, L19501. http://dx.doi.org/10.1029/2004GL020938. Reichle, R.H., Koster, R.D., De Lannoy, G.J., Forman, B.A., Liu, Q., Mahanama, S.P., Touré, A., 2011. Assessment and enhancement of MERRA land surface hydrology estimates. J. Climate 24 (24), 6322–6338. http://dx.doi.org/10.1175/JCLI-D10-05033.1. Reichle, R.H., Liu, Q., 2014. Observation-Corrected Precipitation Estimates in GEOS-5. NASA/TM 2014-104606. . 35. 20150000725. Rienecker, M.M., Suarez, M.J., Gelaro, R., Todling, R., Bacmeister, J., Liu, E., Bosilovich, M.G., Schubert, S.D., Takacs, L., Kim, G.K., Bloom, S., 2011. MERRA: NASA’s modern-era retrospective analysis for research and applications. J. Climate 24 (14), 3624–3648. http://dx.doi.org/10.1175/JCLI-D-11-00015.1. Rumelhart, D., Chauvin, Y., 1995. Backpropagation: Theory, Architectures, and Applications. Lawrence Erlbaum Associates. Santi, E., Paloscia, S., Pettinato, S., Fontanelli, G., 2016. Application of artificial neural networks for the soil moisture retrieval from active and passive microwave spaceborne sensors. Int. J. Appl. Earth Obs. Geoinf. 48, 61–73. Santi, E., Pettinato, S., Paloscia, S., Pampaloni, P., Macelloni, G., Brogioni, M., 2012. An algorithm for generating soil moisture and snow depth maps from microwave spaceborne radiometers: Hydroalgo. Hydrol. Earth Syst. Sci. 16 (10), 3659–3676.

130

J. Kolassa et al. / Remote Sensing of Environment 191 (2017) 117–130

Schaefer, G.L., Cosh, M.H., Jackson, T.J., 2007. The USDA natural resources conservation service soil climate analysis network (SCAN). J. Atmos. Ocean. Technol. 24, 2073– 2077. http://dx.doi.org/10.1175/2007JTECHA930.1. Schmugge, T.J., O’Neill, P.E., Wang, J.R., 1986. Passive microwave soil moisture research. IEEE Trans. Geosci. Remote Sens. GE-24, 1. http://dx.doi.org/10.1109/ TGRS.1986.289584. Seneviratne, S.I., Lüthi, D., Litschi, M., Schär, C., 2006. Land–atmosphere coupling and climate change in Europe. Nature 443 (7108), 205–209. http://dx.doi.org/10. 1038/nature05095. Sevanto, S., McDowell, N.G., Dickman, L.T., Pangle, R., Pockman, W.T., 2014. How do trees die? A test of the hydraulic failure and carbon starvation hypotheses. Plant Cell Environ. 37 (1), 153–161. http://dx.doi.org/10.1111/pce.12141.

Troch, P.A., Vandersteene, F., Su, Z., Hoeben, R., Wuethrich, M., 1996. Estimating Microwave Observation Depth in Bare Soil through Multi-Frequency Scatterometry. 1st EMSL User Workshop Proceedings, Ispra, Italy. Wagner, W., Hahn, S., Kidd, R., Melzer, T., Bartalis, Z., Hasenauer, S., Rubel, F., 2013. A review of its specifications, validation results, and emerging applications. Meteorol. Z. 22 (1), 5–33. http://dx.doi.org/10.1127/0941-2948/2013/0399. Wagner, W., Lemoine, G., Rott, H., 1999. A method for estimating soil moisture from ERS scatterometer and soil data. Remote Sens. Environ. 70 (2), 191–207. http:// dx.doi.org/10.1016/S0034-4257(99)00036-X.