1. Introduction
Surface soil moisture (SM) is a key component of the Earth system which can impact weather through its influence on evaporation and surface energy fluxes [
1], with a lack of SM associated with drought occurrence [
2] and an excess related to flooding [
3,
4]. Given its importance for regional-to-global hydroclimate variability and change, SM is considered as one of the 50 Essential Climate Variables (ECVs) in the Climate Change Initiative (CCI) project established by the European Space Agency (ESA) [
5]. In a Canadian context, SM availability over agricultural regions has been identified as a limiting factor to agricultural yields [
6]. Furthermore, the Prairie region, which accounts for more than 80% of agricultural land in Canada, is an important global food supplier [
6,
7], thus highlighting the need for SM measurement for applications related to agriculture in Canada.
SM is difficult to measure due to its inherent spatial and temporal heterogeneity [
8], and even though there are multiple methods available to measure SM, they all have limitations. Ground-based monitoring stations provide accurate point-scale SM estimates but they usually lack the spatial coverage to capture SM conditions over larger regions and these estimates may not be representative of soil conditions even within the vicinity of these stations [
9,
10,
11]. Another common approach is to use microwave remote sensing of SM, which is promising due to the almost linear relationship between the microwave radiance emitted or reflected by the surface soil portion and the soil-water mixing ratio [
12]. Even though microwave remote sensing of SM can also be challenging due to reasons related to topography [
13], the presence of snow and ice [
12], human-induced radio frequency interference (RFI) [
14] and vegetation water [
15], it provides a means to overcome the spatial limitations posed by in situ observations [
16,
17].
The Soil Moisture and Ocean Salinity (SMOS) mission [
18,
19], launched by the European Space Agency (ESA) in November 2009, was the first Earth observation mission dedicated to SM mapping. The SMOS satellite uses L-band technology (1–2 GHz), which is known to minimize attenuation of microwave signals by the atmosphere and underlying vegetation [
18,
19,
20,
21,
22,
23] and was shown to capture SM over time in a stable and consistent pattern over Canadian agricultural regions [
24,
25]. However, SMOS is also known to exhibit a dry bias, especially over arid regions [
25,
26,
27,
28,
29,
30] and to overestimate SM after large rainfall events [
25,
29]. There have been attempts to improve SM retrievals from SMOS, such as the development of a gridded multi-orbit (MO) SMOS Level 3 product [
31] which showed improvements in the number of successful retrievals and wetter retrievals as compared to SMOS Level 2. SMOS-INRA-CESBIO (SMOS-IC) [
32] is another more recent product which was shown to correspond better to the European Centre for Medium-Range Weather Forecasts (ECMWF) SM as compared to SMOS Level 3 [
32].
Combining information derived from satellite-based passive and active microwave sensors has been reported to, potentially, not only provide improved SM estimates globally but also better spatial coverage and increased number of observations [
33]. This approach of blending multiple products has also been employed to improve estimates of other climate variables, such as snow water equivalent [
34]. The ESA CCI (Climate Change Initiative) SM project was established to develop a long-term soil moisture product from multiple active and passive microwave sensors, in response to the need for a climate data record of satellite SM [
35]. This blended active-passive product combines the strengths of each included single-sensor SM product, one of which is SMOS [
35]. Given its potential to provide improved SM estimates and the need for SM measurement over Canadian agricultural regions [
6], the ESA CCI product could therefore be a useful tool for agricultural applications over Canada.
While the ESA CCI product has been evaluated against ground observations over the United States, China and Europe [
36,
37,
38,
39,
40,
41], such studies over Canada are limited [
42,
43]. Furthermore, the impact of algorithmic changes to the merging of active and passive satellite products in the last major release of ESA CCI SM [
44] on SM retrieval over Canada is still unknown. To date, the assessment of SMOS Level 3 products over Canada is limited to its sensitivity to freeze/thaw cycles [
45], while SMOS-IC has not been evaluated to our knowledge. Therefore, the primary objective of this study is to evaluate whether the recent improvements to the SMOS and ESA CCI products have led to improved accuracy in the SM retrievals over Canada, compared to in situ observations. A secondary objective is to compare the best performing versions of SMOS and ESA CCI and determine which one is best overall for agricultural applications over Canada.
This article is organised as follows.
Section 2 contains a description of the in situ networks and satellite-based products used in this study and the methods used to evaluate the datasets. We present the results of the comparison in
Section 3. In
Section 4, we discuss the accuracy of the individual satellite SM products and the key differences between them.
4. Discussion and Conclusions
Two versions of the ESA CCI SM product, v3.3 (CCI3) and v4.2 (CCI4), SMOS-L3 (L3) and a more recent SMOS product known as SMOS-IC (IC), were evaluated against regional (Ontario and Manitoba) and provincial (Alberta) in situ daily SM monitoring networks over important agricultural regions of Canada. We found that SMOS products generally show higher temporal correlations with in situ measurements as compared to ESA CCI, irrespective of the soil texture or location. However, SMOS products tend to have larger biases, RMSD and SM variability than ESA CCI but are able to better capture anomalies, even though all products capture drying better than wetting.
Overall, ESA CCI could be more appropriate than SMOS for short period studies spanning over a few days or weeks, especially at regional scales, due to the large difference in the number of observations between them. However, given their ability to capture anomalies and SM variability at regional and provincial scales, for periods during which there are sufficient observations, SMOS is more suitable for agricultural applications over Canada than ESA CCI. We also found that, overall, CCI4 and IC outperformed their counterparts in all comparisons.
Several possible reasons have been reported for the dry bias in SMOS [
26,
41]. First, the sampling depth for SMOS at 0–3 cm is shallower than the in situ SM measurement at 5 cm [
70]. Second, it is possible that the SM measured in situ could be overestimated due to soil compaction, especially during dry periods [
71]. Third, Cui et al. [
41] reported that an underestimation in surface temperature could be a factor causing the dry bias in SMOS. Finally, RFIs can increase the recorded TB in SMOS and thus a dry bias in the retrieved SM [
14]. Furthermore, the drier bias in IC as compared to L3 is consistent with previous literature [
72] and could be due to the small constant initial value for SM in the TB cost function in IC [
32]. As for ESA CCI, in case RFIs are detected in multi-frequency retrievals, retrievals at a higher frequency such as X-band are selected [
35], which could explain the smaller bias as compared to SMOS, even though the bias in ESA CCI is also characteristic of the bias present in GLDAS-Noah since its dynamic range is imposed on ESA CCI [
73,
74].
Our results also suggest that the “median of metrics” method is more appropriate than the more common “metrics of mean SM” method for evaluating the correlation and variability of satellite products with large in situ networks (
Figure 4). While the small spatial variability for Ontario shows good representativeness of all stations, the large spatial variability over Manitoba due to high soil diversity shows weak representativeness of some stations over that network and agrees with previous literature [
11,
25]. However, while the low correlation values over clay soils did not have a large influence on the overall correlation for Manitoba, the observed lower SM variability of these soils can influence the comparison of trends. For instance, clay soils can retain SM longer than sandy soils due to having lower SM variability [
25]. Since all satellite products showed considerable drying when the stations in Manitoba showed wetting, it could mean that the observed wetting for that network is an artefact resulting from low SM variability at the clay stations.
In
Section 3.1 we showed that of the SMOS products, L3 has much higher temporal variability than IC over Ontario, as compared to Manitoba and Alberta. We speculate that L3 could be influenced by the relatively higher spatial variability in land cover over Ontario, which is not present over Manitoba and Alberta; for example, the individual satellite pixels that include the Ontario stations contain large fractions (>50%) of forest, open water, and non-agricultural land [
25]. The L3 retrieval algorithm takes into account pixel land use and heterogeneity while IC does not, and therefore could be impacted by the uncertainties present in the auxiliary datasets used to characterise pixel heterogeneity [
32]. We believe that other methodological differences in the IC product, such as the use of a more robust TB product, and the updated L-MEB vegetation and soil parameters [
32], are less likely to explain the observed differences in SM variability. The effect of changing the TB data removes outlier retrievals that were present in L3, reducing the overall number of non-missing observations in IC; however, we compare products only for days when data is available from all products. Furthermore, updated land cover parameters in IC differ from L3 only for areas of low vegetation, while for forested areas the parameters are the same in both products [
32].
This study attempted a pixel-to-point comparison, which presents well-known challenges due to inconsistencies in the spatial averaging scale of satellite products versus in situ probe measurements and the treatment of pixel heterogeneity in satellite products. Future work including spatial gridded comparison between these satellite products and reanalysis products, such as the recently released fifth generation of ECMWF reanalysis, ERA5 [
75], could reveal more information about the spatial characteristics of these products and the consistency among these different products. Finally, we note that a more recent L-band product, the Soil Moisture Active and Passive (SMAP) [
20], launched in 2015, has been shown to capture relative soil moisture trends well over Canada [
76]. We anticipate that future work comparing IC and SMAP for a more recent period should evaluate which of these products performs best for monitoring SM variability across Canadian agricultural regions.