1. Introduction
Seasonal freeze/thaw (FT) impacts about half of the northern hemisphere [
1]. It is a dominant control on the water, energy, and carbon cycle, including groundwater and surface water dynamics; exchange of latent and sensible heat controlled by vegetation; and snow and soil processes [
2,
3,
4,
5,
6,
7]. While the impacts of FT have been studied in great depth at boreal and higher latitudes [
8], there are also examples of impacts on hydrological processes [
9,
10] and roads [
11] in the contiguous U.S. (CONUS).
A lack of in situ soil temperature observations presents a key knowledge gap in assessing frozen soil extents. Owing to limited in situ soil observations, seasonal FT studies are often strictly limited to observed or modelled air temperatures [
4]. However, air and soil temperatures will usually be different as snow, vegetation, litter, and organic layers insulate soils. Soils may be relatively warmer (colder) than air temperature during freeze-up (thaw), or they may not freeze at all.
There is mounting evidence that passive microwave FT observations provide a transformative means to improve our understanding of the spatiotemporal FT processes for a variety of landscapes [
1,
4,
12,
13,
14]. Typically, the retrieval of the FT state from passive microwave observations uses a change detection approach to identify changes to the dielectric constant using a brightness temperature threshold, a moving average window, or edge detection [
4]. These approaches have been successfully applied for almost two decades primarily using 19 and 37 GHz observations from SSM/I (and to a lesser extent, from AMSR-E). More recently, studies have successfully used L-band (1–2 GHz) observations from space to detect FT state primarily via the Soil Moisture and Ocean Salinity (SMOS) [
13,
14,
15,
16,
17], SAC/D Aquarius [
18], and Soil Moisture Active Passive (SMAP) observations. L-band is more effective at detecting soil FT as compared to higher frequencies. L-band corresponds to a greater emission depth and is less impacted by vegetation [
13,
14,
15].
A current omission of recent L-band FT studies is that they did not extend to latitudes below 45°N. Impacts of seasonally freezing soils are not limited to just these northern regions. Earlier work using passive microwave observations at higher frequencies included regions below 45°N. For example, Zhang et al. (2003) developed a FT algorithm for CONUS using the 19 and 37 GHz bands of SSM/I [
1]. Their maps suggested that frozen ground would be found in most of CONUS. Kim et al. (2017) used SSMR and SSM/I data (37 GHz) to generate global frozen landscape extents based on data ranging from 1970–2008 [
18]. Those results showed that most of North America froze at one point during winter. Therefore, L-band FT retrievals should yield good results in at least some regions south of 45°N.
This study evaluates FT retrievals at SMAP core validation sites (CVS) located in CONUS. SMAP CVSs are densely sampled and usually consist of about ten stations covering a spatial extent of about 40 km by 40 km. Previous studies have used boreal (>45°N) latitude CVSs to validate the FT product [
19,
20]. Sites below 45°N have been used for soil moisture validation [
21], but some of these sites should also be well-suited to assess SMAP FT retrievals at mid-latitudes.
We use a detection approach similar to that used for the NASA SMAP FT product (L3_FT_P). Our hypothesis is that SMAP landscape FT retrievals would often (e.g., >70%) correspond to soil FT states at mid-latitude CVSs. This study is primarily concerned with evaluating SMAP retrievals against 0–5 cm soil temperature data collected at mid-latitude CVSs, but also maps annual (2016–2018) freeze extents in CONUS.
5. Discussion
Derksen et al. (2017) found that the correlation between NPR time series with in situ temperature was improved for AM (~0.8) over PM (~0.3) observations. That study also found that SMAP PM FT data agreed better with in situ observations. This study found somewhat improved correlations for PM data in Iowa and Indiana: FT flag agreement improved about 5% when PM data was used. It is possible that PM validation metrics improved in part due to higher uncertainty in AM data caused by refreezing [
19]. It is difficult to make a statement on error bias between AM and PM in this study, because commission or omission errors were similar (within 5%). Additionally, Derksen et al. (2017) limited their study to a thawing landscape, because it focused on the period during which the active radar collected data (April–July 2015) [
19]. Our FT study focused on winter (October–March) but also included months during which landscape freeze-up and thaw occurred in CONUS.
Although only a few sites were investigated, it was noted that validation metrics improved where greater ΔNPR values were found (
Table 3 and
Table 4). NPR
fr was nearly constant between AM and PM overpasses, whereas NPR
th was about 20% greater (
Table 3,
Figure 3,
Figure 4 and
Figure 5). Overall accuracy for PM data was usually about 5% greater compared to AM. A likely explanation is that STA can, to a certain extent, more confidently delineate between freeze and thaw if greater ΔNPR values are used as input to the algorithm. Thus, the current approach for setting NPR references may produce a smaller than optimal ΔNPR range to be used in conjunction with STA, and accuracy metrics (for Δ(t)
thr = 0.5) are not as good as they could be. It is also important to keep in mind that this work did not include any error mitigation efforts, and accuracy metrics may be improved further.
To follow up on the question as to why NPRth(PM) values could be 10–20% greater than NPRth(AM), a small case study was made for one Idaho grid (60901). Dates corresponding to the maximum five NPRth were 11 July 2015, 13 July 2015, 10 July 2016, 12 July 2016, and 14 August 2017 (11 July 2015, 14 July 2015, 11 July 2016, 27 July 2017, and 15 August 2017) for AM (PM) data. Dates obtained for AM and PM are nearly identical: Except for one “pair”, they are no more than one day apart from one another. The meteorological record at a weather station in Boise, ID (~50 km distance) indicated that there were only two precipitation events >1 mm in July/August 2015/2016, namely, on 10 July 2015 between 8 AM—4 PM (about 8 mm) and on 10 July 2016 between noon and 6 PM (about 15 mm). It follows that if it is true that summer convective precipitation would provide more water to soils during a 6 AM to 6 PM window vs. a 6 PM to 6 AM window, then it should be expected that NPRth(PM) would be somewhat increased for 6 PM observations relative to those made at 6 AM—and explain the increase in NPRth(PM).
An important limitation of methodology is that SMAP FT detection relies on a method traditionally used to estimate water content in landscape elements (vegetation, and soils). NPR is equivalent to the Microwave Polarization Difference Index (MPDI), which is a well-established quantity that has been used to characterize whether a landscape (e.g., soil, vegetation) is wet or dry [
35,
36]. Liquid water, dry soil, and ice, respectively, have dielectric constants (ε
r) of approximately 80, 5, and 3 [
3,
37]. Therefore, frozen soil—irrespective of its frozen water content—would have a dielectric constant that is comparable to that of dry soil. Thus, SMAP classification results should then also be interpreted as ‘wet’ (thawed) and ‘dry’ (frozen).
The higher incidence of SMAP ‘frozen’ retrievals during summer (
Figure 4) can be attributed to Idaho being relatively drier than the other CVS (semi-arid,
Table 1). For dry sites such as the Idaho CVS, it would be valuable to develop and test alternative FT retrieval methods that do not depend on NPR, such as the one presented in Kim et al. (2011) [
4]. That method also used the STA approach, but applied it to Tb
V rather than NPR. L3_FT_P version 2 uses the Tb
V-based type of FT retrieval at most grids below 45°N [
38].
Accuracy assessment of SMAP landscape FT retrievals is difficult due to the impact of landscape heterogeneity on coarse resolution observations. The SMAP radiometer has an ellipsoidal instantaneous field of view of 38 by 49 km and therefore incorporates landscape elements that are not accounted for with station point data. At the studied CVSs, there are geographic biases in the siting of stations with respect to each SMAP grid: Stations in grids 60901 (Idaho), 62891 (Iowa), 62892 (Iowa), and 65806 (Indiana) are located inside a radius of about 20 km. Stations in grids 68165, 63855, and 63856 are located inside a radius of 10 km or less, and they cover less of the SMAP grid area.
The best accuracy metrics between SMAP FT and in situ temperature data were obtained in Iowa (grid 62891) and Indiana (grid 65806). Relatively poorer results in Idaho may be attributed to the small ΔNPR and the semi-arid climate. The semi-arid climate factors into a higher rate of errors of commission, because NPR/MPDI discriminates between dry/wet rather than freeze/thaw.
Idaho also features a heterogeneous landscape that is better represented by in situ data if all station data are used, rather than dividing station data according to grids (
Section 4.3). Compared to the boreal study, ΔNPR at grasslands and croplands situated in CONUS were substantially smaller. In the boreal study, ΔNPR averaged about 2.9 (3.9) for grassland (cropland) compared to our results of 1.5 (2.6). While lower, ΔNPR obtained at CONUS CVSs are within one standard deviation of those obtained in the boreal NASA SMAP FT study. Even at fairly homogenous CVSs (Iowa and Indiana), significant sub-grid variability exists (
Figure 9).
The relatively greater ΔNPR obtained in the boreal study for grasslands and croplands could potentially be attributed to NPR
fr values in that region that are more representative of fully frozen soils than the current study. Reliable NPR
fr values need soils within a landscape (SMAP grid) to be frozen to a greater depth than the L-band penetration depth—a condition that is more difficult to satisfy at lower latitudes. However, data obtained in the boreal study for the Canadian Prairie region (grasslands and croplands) do not support this idea: The prairie NPR
fr values (~3) exceeded those at SMAP CVSs (2.4). However, prairie NPR
th values were much larger (>6) than those obtained at SMAP CVSs (~5). The boreal study’s ΔNPR map indicates that the Canadian Prairie region typically has ΔNPR values between 3 and 4, a similar range to the mean values that were reported for their study’s grasslands (2.9) and croplands (3.9). Thus, the relatively smaller ΔNPR for CONUS grassland and cropland sites could more likely be attributed to their relatively smaller NPR
th. While the Canadian Prairie region is dry, the region’s climatology indicates that much of their annual rainfall totals occurs during a July/August window, with July/August totals decreasing with decreasing latitude [
39]. If the 2015 July/August precipitation totals followed a similar spatial pattern, then this could explain why the July/August NPR
th values were greater in the boreal study than for CONUS. However, this perspective is limited to only looking at the Canadian Prairies and a few locations in CONUS. Further exploration is needed to conclusively support the above conjecture regarding greater NPR
th and ΔNPR mean values for grasslands and croplands in the boreal study compared to at CONUS CVSs.
Another important aspect is that landscape heterogeneity can also be caused by ephemeral water (EW). Because FT algorithms exploit the substantial dielectric constant differences as water transitions from frozen to liquid (and vice versa), it is important to be aware that the SMAP observations would be additionally impacted by EW on land. While SMAP Tb values are routinely corrected for static water [
30], EW is a potential source of error. For example, wet snow may result in a ‘thawed’ SMAP FT retrieval, while the soil may still be frozen [
19]. In this case, the soil state cannot be directly detected, because a surficial water layer or wet snow masks the soil’s emission. Spring and midwinter thawing may also produce ephemerally flooded areas within a landscape. Due to its high sensitivity to liquid water, even a localized event can significantly impact the SMAP FT retrieval accuracy within its footprint. However, dry snow can also be a source of error in FT retrievals at L-band because of refraction and impedance matching [
40,
41]. These effects impact Tb such that the Tb signal more closely corresponds to that of a frozen soil: The NPR of frozen soil covered by dry snow would be lower than that of bare frozen soil. Thus, it is possible to have a false frozen retrieval in the case of a wet soil being covered by dry snow. This situation is possible early in the cold season. If snowpack properties and moisture at the snow/soil interface were available, we could study this potential source of error in more detail.
Limiting assumptions related to FT to binary classification can also impact accuracy assessment, mainly because relatively small differences in temperature change the classification of soil FT. Small measurement uncertainty or instrumentation bias could lead to errors in observed soil state. Thermistors used at the CVSs are optimally calibrated for a temperature of 20 °C and would have errors ranging from ±0.1–0.3 °C at near-freezing temperatures. In situ FT detections might be more robust if the soil dielectric constant were used instead of soil temperature. Also, the temperature at which natural soils freeze is usually lower than 0 °C, and a significant portion of water may remain in a liquid phase until soil temperatures fall well below freezing (e.g., −0.5 °C). Freezing point depression would impact SMAP FT validation accuracy metrics, because soil may still be wet/thawed when its physical temperature is less than 0 °C. During thaw, soils often become isothermal for an extended period of time. In this state, both ice and water are present, and it is reasonable to refer to this state as either frozen or thawed.
It is possible to improve on the shortcomings of binary classification by aggregating in situ soil temperature data according to SMAP FT retrieval. In this representation, if SMAP retrievals are sensible, frozen soils would be colder and show median temperatures close to or below freezing.
Figure 7 showed that SMAP landscape FT retrievals corresponded quite well with soil temperatures at most grids, indicating that there should be good confidence in SMAP FT retrievals of the landscape state corresponding to soil temperature at least at some locations in CONUS.
A priori knowledge of where SMAP FT retrievals are accurate would be especially valuable for sub-boreal latitudes, because it is important to only define freezing thresholds where landscape elements (i.e., soils) freeze. Inappropriate application of thresholds will cause the interpretation of results to be extremely difficult, because classification results are impacted by retrievals over forested areas or in climates that are both dry and cold. Version 2 of L3_FT_P addresses this issue by only using the STA at those locations where model data indicated frozen conditions for at least 20 days per year. This work also identified some indicators leading to improved SMAP FT retrievals. Here, the optimal frozen condition appears to be greater than 20 days. The best performance was obtained in the Iowa CVS. Iowa had nearly twice the frozen duration per year (75 days) and colder soil temperatures that were considerably colder than Idaho or Indiana. The temperature of frozen soils should also be considered. SMAP landscape FT retrievals would probably be more accurate if soils froze longer and colder (e.g., average temperature below <−1.0 °C). Finally, the STA is not recommended for locations where ΔNPR < 2.