*3.2. Comparison of SMOS, GLDAS-Noah, ERA5 and In-Situ at Target Sites*

Time series of SMOS-based SM together with top 10 cm soil moisture GLDAS-Noah and top 7 cm soil moisture analysis from ERA5 are shown in Figure 8. It can be observed that in regions C to H, SMOS, GLDAS and ERA5 soil moisture time series are comparable in terms of temporal phase (Pearson correlation of 0.7–0.9). In regions A and B, however, the SMOS observations are uncorrelated with the uppermost soil moisture modeled estimates. This is mostly due to the presence of snow, which masks satellite measurements and can also affect the uncertainty of model predictions. Further research is needed in these areas of mutual disagreement to identify potential deficiencies in the satellite and/or in the model estimates. Similar results were reported by a previous study comparing SMOS with MERRA reanalysis [43]. Note that the correlation metric benefits from the seasonal cycle, which for the period of this study is included. Correlation of time series at daily time scales lead to similar results (not shown).

There are notable differences in the absolute values of satellite and modeled data (e.g., sites E and H), which remarks the need to use bias correction procedures when assimilating soil moisture satellite observations in land surface simulations [44]. The assimilation of satellite data in land surface models has proven to be a powerful technique to leverage from the two sources of information. A recent study has shown that assimilating SMOS SM data into the Noah Land Information System after bias correction significantly increased the anomaly correlation of modeled top soil moisture estimates with station measurements [45]. Also, Pinnington et al. [46] showed that assimilation of (bias-corrected) satellite rainfall and SM data had the greatest impact on model estimates during the seasonal wetting-up and drying-down of the soil, respectively. While L-band satellites have been designed for measuring soil moisture, land surface models have been designed for a much wider purpose, including ecological, hydrological or climate applications. Modeled soil moisture is generally highly sensitive to the meteorological forcing data used and the land surface model encoded physics [47,48]. This makes comparison of absolute values of observed and modeled SM

very challenging, and therefore studies generally focus on the comparison of SM temporal anomalies (e.g., [19,20]), or even in the comparison of observed and modeled brightness temperatures directly (e.g., [49]). The differences found between modeled and observed SM and also among different models support the idea proposed here of leveraging from the natural soil moisture variability captured by satellite observations as a reference for harmonizing SM climate data records so that they become model-independent.

**Figure 7.** STL decomposition of the time series at the eight target locations (from Figure 1 and Table 1). The smoothed SMOS signal (in black) is decomposed in its seasonal cycle (in blue), linear trend (dashed green line), interannual variability (green), and residual component (red).

Ground-based soil moisture data from REMEDHUS (average of 17 in-situ stations) are also included in Europe time series (Figure 8E). It can be seen that both satellite and models generally capture the temporal dynamics measured by the in-situ sensors. A statistical analysis has been

undertaken following the recommended performance metrics in [50]. Results are shown in Table 2. The temporal correlation (R) and the unbiased Root-Mean-Squared difference (ubRMSD) between the in-situ, remotely sensed and modeled SM are satisfactory (R > 0.85, ubRMSD < 0.003). As in the previous analyses, it should be noted that the R metric benefits from the seasonal cycle, which for the period of this study is included. In terms of accuracy, SMOS shows a low dry bias of 0.02 while the two models present a large wet bias of about 0.11 m3·m−3. As expected, the models capture reasonably well the temporal dynamics but not the absolute magnitude [47]. The statistical scores when calculated at daily time scales lead to similar results (not shown).

**Figure 8.** Time series of SMOS (in black), top 10 cm GLDAS-Noah (in green) and top 7 cm ERA5 (in blue) soil moisture at the eight target locations (from Figure 1 and Table 1). Target location E also includes time series of collocated ground-based soil moisture (dashed line in magenta, average of 17 REMEDHUS in-situ stations).

**Table 2.** Statistical scores from comparison to in-situ at REMEDHUS (time series on Figure 8E): bias, unbiased Root-Mean-Squared difference (ubRMSD) and Pearson correlation (R).


#### *3.3. Temporal Mean and Variance of SMOS Soil Moisture Retrievals at the Global Scale*

The temporal average global map of SMOS SM surface volumetric soil water content for the study period (pixel average of *SMtot* in Equation (1)) is shown in Figure 9. The average conditions observed during the study period exhibit a mode of 0.1 m3·m−<sup>3</sup> (inset figure). It shows the expected spatial patterns of SM, from the dry arid regions to the wet forested areas. Note that the average conditions shown in areas with limited availability of data may not be representative (e.g., northern latitudes and tropical Africa, see coverage map on Figure 4).

**Figure 9.** Global distribution of time-average soil moisture based on six years of SMOS observations, starting in 1 June 2010. The inset figure is an estimate of its marginal probability density.

The global map of SMOS SM standard deviation for the study period (standard deviation of *SMtot* in Equation (1)) is shown in Figure 10. As expected, areas with higher variability are concentrated in the tropics, where there is a strong seasonality dominated by the position of the ITCZ [51]. Other areas with strong variability include India and South-East Asia, strongly affected by Monsoon rainfall, and some regions in the Southern Hemisphere like eastern Australia and South America that could be related to ENSO. In particular, the high SM variability observed in the so-called Southeastern South America (SESA) region can be explained by the intense summer precipitation over this region [52]. This pattern has also been observed with satellite SM products derived from higher microwave frequencies and climate models [53], and responds to strong land-atmosphere interactions in the region. The variability observed in northern latitudes is probably due to imperfect detection of ice/snow and poor temporal coverage (see Figure 4).

#### *3.4. Analysis of the Dominant Features of Global Soil Moisture Variability*

In this study we decomposed the total variance of the soil moisture signal (Figure 10) by breaking it down into: (1) a seasonal component, (2) a long-term component, and (3) a high-frequency residual or subseasonal component. The relative magnitude of each of these three components is shown as an RGB triplet in Figure 11. It reveals the seasonal cycle is dominant in many tropical regions such as Brazil, Central Africa, India and Northern Australia. Wet tropical forests such as the Amazon and Congo exhibit high spatial heterogeneity in its dominant components, probably due to the combined effects of human activities and climate variability. In contrast, the soil moisture variability in dry tropical forests, including the savannahs south of Central Africa and several regions in Southwestern Brazil, is clearly dominated by seasons. Though dry tropical forests may receive several hundred centimeters of rain per year, they have long dry seasons which last several months and vary with geographic location. Forests in dry tropics show great diversity of phenological patterns and large interannual variation, with predominance of deciduous tree species [54]. Our analysis indicates that L-band is capturing the emissivity from the soil and the seasonal drought through the forested canopy. Two regimes can be identified in Europe: whereas western countries are governed by seasonal variability, long-term variability predominates in the eastern countries. In Australia, seasonality dominates the north and the south-east, the eastern region is dominated by long-term variability and the western by subseasonal. The Indo-Australian archipelago is also dominated by long-term variability. It is interesting to note that subseasonal variability is predominant in regions where the SMOS signal has already a relatively low variance (Figure 10) and is most likely influenced by noise such as the Sahara desert, the Arabian Peninsula and Western Australia, i.e., where the noise is at least of the same magnitude that the annual variance. Indeed, despite having carefully filtered the data, some of the SMOS retrievals in Asia and Europe may still be affected by undetected RFI contamination [27]. Also, results may be affected by the intrinsic uncertainty of microwave satellite SM retrievals, which is higher in presence of dense vegetation canopies, heterogeneous landscapes, and high topography [6,11,55].

**Figure 10.** Global map of SMOS soil moisture standard deviation for the six-year period of this study. The inset figure is an estimate of its marginal probability density.

The linear trends within the long-term variability component are further examined in Figure 12. The linear trends observed in the SMOS signal illustrate the ENSO conditions during the six-year study period. During the first five annual cycles -from 2010 to 2015- the equatorial Pacific Ocean was mostly in a cold phase (La Niña); however, a warm event (El Niño) occurred during the last cycle (i.e., in 2016). This explains why SMOS-derived linear trends reproduce known ENSO teleconnection spatial patterns. During El Niño, limitations in terrestrial moisture supply result in vegetation water stress and reduced evaporation in eastern and central Australia, southern Africa and Eastern South America (areas in red). The contrary situation is experienced in Argentina, Tanzania and southeastern US (areas in blue), where there is a regime of above-average rainfall. This result is in line with a previous study that showed that multi-decadal (1980–2011) variability in SM and terrestrial evaporation was dominated by ENSO dynamics [56]. The linear trends in Figure 12 also provide evidence of strong dry/wet patterns in regions which have not been previously related to ENSO precipitation patterns (e.g., western Europe, Northern Africa, California). Still, results should be taken with caution, and further analysis of the

long-term variability is needed to identify new climate patterns at short and median temporal scales. Investigating the relation of ENSO to global soil moisture variability and the existence of potential new teleconnection patterns is recommended for future research.

**Figure 11.** Distribution of the total SMOS variance among the long-term (green), seasonal (blue) and subseasonal (red) components, expressed in per cent of the total variance, indicating the dominant models of temporal variability in soil moisture for different regions. Each vertex in the triangle corresponds to 100%. Empty areas correspond to locations with less than 80% SMOS temporal coverage that have been masked out.

**Figure 12.** Magnitude of linear trends in the SMOS signal (expressed in m3·m−3·year<sup>−</sup>1). The trends reflect that during the SMOS period (2010–2016) the equatorial Pacific Ocean was mostly in a cold phase (La Niña) with a transition to a warm phase (El Niño) in 2015–2016.

#### **4. Discussion and Final Remarks**

The two first space missions dedicated to measuring the Earth's surface soil moisture have been launched in the last decade: SMOS in 2009 and SMAP in 2015. They are providing L-band measurements that, combined with C- and X-band measurements available since the 1980s, allow generating, for the first time, observational soil moisture climate data records combining the three microwave frequency bands. However, the synergies of microwave measurements across different frequencies and the potential of the L-band data record to serve as a common reference to harmonize the long-term data record have not yet been fully exploited. This study presents first evidence that the relatively short data record of available SMOS L-band observations allows characterizing the main modes and spatial distribution of the Earth's surface soil moisture variability. This information could serve as a reference to harmonize and construct a model-independent SM climate data record. In its current version, the ESA CCI soil moisture product uses a climatology obtained from GLDAS-Noah to harmonize the individual products for the 40-year period of record. However, given the differences found between modeled and observed SM, and among SM from different models, there is a clear need for a SM climate data record based solely on observational data sets. Such a record will be instrumental to verify land surface model performance and trends.

This study is a first attempt to derive a global L-band climatology from observational data alone. The STL geostatistical procedure was implemented and tailored to the first six annual cycles of SMOS data to decompose the temporal variability of the signal. For the period of study (2010–2016) this analysis allowed identifying regions where soil moisture variability was dominated by seasonal cycles, regions that did not exhibit a clear seasonal pattern and with likely subseasonal variability, and regions where the long-term variability dominated. Results show that the seasonal cycle was dominant in the tropics (Brazil, Central Africa, India and Northern Australia), and in dry tropical forests, including the savannahs south of Central Africa and several regions in Southwestern Brazil. Wet tropical forests, in turn, exhibited high spatial heterogeneity in its dominant components, probably due to the combined effects of human activities and climate variability. In Europe, western countries were governed by seasonal variability whereas the long-term variability predominated in the eastern countries. In Australia, seasonality dominated the north and the south-east and the eastern region was dominated by long-term variability. Interestingly, in regions where SMOS has very limited variance (e.g., Sahara desert, Arabian Peninsula, Western Australia), the subseasonal or residual component was dominant. This is probably due to the fact that the noise of the observations is at least of the same magnitude that the annual variance and the dominant variability is identified as noise. During the study period (2010–2016), the equatorial Pacific Ocean was mostly in a cold phase until 2015–2016 when there was a transition to a warm phase. The observed global linear trends, based upon the strong El Niño event in 2016, are shown to be consistent with the ENSO teleconnections calculated over multiple events.

This study has demonstrated that L-band observations provide a reliable source of data to monitor the distribution of shallow water content in continental surfaces from space platforms. However, this study cannot conclude on the presence of a clear climate trend in the water content of the soil and results should be taken with caution, since they are limited by the length of the L-band data record available (∼8 years vs. the 30 years which are generally considered for climate studies). Still, our results are consistent with a previous study that showed that multi-decadal (1980–2011) variability in SM and terrestrial evaporation was dominated by ENSO dynamics [56]. Also, there are relevant studies considering the trends of SM and terrestrial evapotranspiration at scales smaller than 30 years (e.g., 10 years in [57]).

Our results showed that the STL procedure, after adequate parameterization, was a solid means to build a soil moisture climatology based solely on the temporal dynamics of the data. Yet, it would be recommendable to assess in future research the additional benefits of using techniques exploiting the temporal and spatial components of the data (e.g., [24,58,59]). The global STL parameterization was thoughtfully chosen after an analysis at 8 selected sites with distinct vegetation seasonality and climatic conditions. However, they do not likely cover each climate on the Earth continents and may not be optimal under specific conditions (e.g., areas where small SMOS variance is dominated by noise). A comparison with modeled GLDAS-Noah and ERA5 soil moisture reanalysis in the selected sites, and with in-situ data at the REMEDHUS site (Europe), provides confidence in the obtained results. The exception is in areas where the presence of satellite data gaps -although limited to 20% of the study period- leads to mutual disagreement of model and satellite estimates. It should be noted that the presence of observational data gaps in the satellite time series can potentially bias the obtained climatology and severely limit the applicability of the method to Northern latitudes, where seasonal snow masks soil emissivity and soil moisture retrievals cannot be performed. In this regard, recent studies on the use of Gaussian process regression techniques to mitigate the effect of missing information in Earth observation data are very promising (e.g., [60]).

This work has shown that the relatively short SMOS data record allows providing insight into the dominant modes of temporal variability in the Earth's surface soil moisture. Also, although previous studies have shown some weaknesses of SMOS retrievals (e.g., [11,27,55]), the good correlation obtained with ERA5 modern reanalysis and GLDAS-Noah indicates that an L-band climatology, such as the one proposed in this study, is a reasonable reference to be used for a climate data record exclusively based on remote sensing data. The presented SMOS-based climatology offers a unique view of recent processes governing freshwater fluxes in the water cycle, and allows observing specific phenomena, as the different variability regimes present within the Amazon and Congo basins and the dominance of the interannual variability in wide regions within Europe, United States, Eastern Australia and South America. It also opens the path to forthcoming studies focused on the analysis of the interannual variability and the impact of ENSO in areas that have not been documented so far.

**Author Contributions:** All authors contributed significantly to this research. The initial idea was conceived by M.P. and J.B.-P.; M.P. carried out the data analysis and led the manuscript writing; J.B.-P. implemented the STL decomposition technique and adapted it to SMOS data; J.M.-S. fully contributed to the design and development of the study and provided the ERA5 data. All the authors contributed to the interpretation of the results, participated in the manuscript and approved it.

**Funding:** This research received no external funding.

**Acknowledgments:** M.P. is supported by a Ramón y Cajal Contract (Spanish Ministry of Science, Innovation and Universities). The authors would like to thank the support received from the ESA CCI Soil Moisture project team.

**Conflicts of Interest:** The authors declare no conflict of interest.
