**1. Introduction**

Soil moisture (SM) is an essential component of the terrestrial hydrological cycle ecosystem [1,2]. In ecology, SM affects the growth and activity of vegetation and microorganisms by controlling the division of water [3], and can also mitigate changes in soil organic carbon content caused due to climate warming [4]. In hydrology, SM is not only an important parameter of the water cycle but is also used to infer surface- and ground-water exchanges [5]. In climatology, SM affects regional climate through changing surface albedo, evapotranspiration intensity, and sensible and latent heat fluxes [6]. In the permafrost environment of the Qinghai–Tibet Plateau (QTP), the SM in the active layer is significantly altered by the seasonal freezing and thawing processes and influences the energy

**Citation:** Li, Z.; Zhao, L.; Wang, L.; Zou, D.; Liu, G.; Hu, G.; Du, E.; Xiao, Y.; Liu, S.; Zhou, H.; et al. Retrieving Soil Moisture in the Permafrost Environment by Sentinel-1/2 Temporal Data on the Qinghai–Tibet Plateau. *Remote Sens.* **2022**, *14*, 5966. https://doi.org/10.3390/rs14235966

Academic Editor: Emanuele Santi

Received: 14 October 2022 Accepted: 21 November 2022 Published: 25 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

exchange between permafrost terrain and the atmosphere [7,8]. The accurate information on the spatial and temporal distribution of SM helps advance hydrological, ecological, and climatological studies in permafrost areas [9].

In the permafrost area of 1.06 × 10<sup>6</sup> km<sup>2</sup> of the QTP [10], it is critical for various scientific studies to obtain accurate spatial distribution data of SM on a large scale. Due to the harsh environment and inconvenience of accessing the QTP, the traditional methods of obtaining SM by sampling measurement and monitoring are limited. Several meteorological and hydrological stations have been deployed over the past few decades, and data scarcity has been filled to a degree [11]. These restricted sites are unevenly distributed on the edge of permafrost areas or on seasonally frozen ground areas, which makes it challenging to study SM with high spatial heterogeneity [6,12]. Remote sensing technology has achieved significant advances in SM monitoring, with its unique large-area observation capabilities, and some remote sensing products and reanalysis data have been produced, such as the fifth generation of the land component of the European Centre for Medium-Range Weather Forecast atmospheric reanalysis (ERA5-Land) [13], the European Space Agency Climate Change Initiative (ESA CCI) [14], and the Noah land surface model driven by Global Land Data Assimilation System (GLDAS-Noah) [15]. Xing et al. evaluated seven SM data products (SMAP, SMOS-IC, ASCAT, ERA5-Land, ESA CCI, LPRM, AMSR2) over the permafrost region of the QTP based on in situ SM measurements and found that the SM data of ESA CCI had the highest accuracy [16]. Due to the lack of adequate measurement data, SM products generated by the model assimilation are significantly biased in the QTP [17]. In addition, the SM product data at a spatial resolution of tens of kilometers are affected by mixed pixels and do not accurately describe the SM distribution. The coefficient of variation (CV) analysis is often used to describe significant patterns in regional mean SM content, with the relationship between CV and SM often showing a hysteresis pattern in spatial variability [18–20]. The spatial–temporal SM variations have never been revealed over the permafrost area in the QTP, where the relationship between the CV and mean SM is similar to that found in other regions is unclear [21,22]. The lack of high-spatial-resolution SM data in the permafrost region of the QTP greatly limits the studies of the spatial–temporal distribution characteristics of SM. Few studies are able to analyze the spatial–temporal variations of SM in permafrost areas on a fine scale.

Synthetic Aperture Radar (SAR) has proven its high potential for retrieving highspatial-resolution SM using the backscatter coefficient ( *σ*◦) [23,24]. Some empirical [25,26], semi-empirical [27], and physical scattering models [28] have been developed to relate backscattering with surface SM, roughness, and vegetation. Surface scattering models are commonly used for bare soil including the semi-empirical Dubios model [29], the Oh model [30], and the physically based Integrate Equation Model (IEM) [31]. These are usually integrated with vegetation scattering models, such as semi-empirical water cloud models (WCM), to predict scattering from vegetated areas [32,33]. Studies have developed algorithms for retrieving SM based on IEM or WCM inversion using minimization, LUT, and machine learning approaches. In 2014, He et al. estimated SM in the alpine meadow region by coupling the IEM and the WCM with R<sup>2</sup> and RMSE reaching 0.71 and 0.03 m3/m3, respectively [34]. In 2017, Bai et al. first estimated SM in the alpine steppe region of Magu using Sentinel-1 (S1) data with the WCM [35]. In 2021, Yang et al. coupled the improved Oh model in 2004 and WCM to estimate SM in the Nagqu region based on S1 data and MODIS optical data with the assumption of constant surface roughness [36]. Despite the significant advance in scattering modeling, SM inversion from these microwave scattering models are commonly ill-posed and complicated [37,38]. Besides utilizing microwave scattering, several algorithms have been successively developed and widely used to retrieve SM, such as change detection (CD) and neural network (NN) [39–41]. NN are mathematical models that are commonly trained by vegetation coefficients, backscatter coefficients, and other parameters in studies of SM retrieval using SAR data, which have high requirements for the data volume. However, the NN method requires a large amount of data for training and validation, which greatly limits its application in the permafrost region of the QTP.

The principle of the CD algorithm is based on the assumption that by differencing the backscattering coefficients in two periods, the effect of soil roughness and vegetation is reduced, and the backscatter difference is mainly due to the changes in SM [39]. Thus, it is more suitable than the NN algorithm for permafrost areas where a large amount of training data acquirement is difficult. Gao et al. mapped SM distributions at a 100 m spatial resolution located in Urgell using a CD algorithm by combining S1 and Sentinel-2 (S2) data [40]. Bauer-Marschallinger et al. used the CD algorithm to build the first global SM dataset with 1 km spatial resolution, which greatly advanced the progress of SM studies [42]. Zhu et al. proposed an unsupervised CD method as a pre-processing procedure for multi-temporal retrieval and improved the accuracy of the CD algorithm by reducing the uncertainty caused by changes in vegetation and roughness [43,44].

Over the QTP, studies of SM retrieval using SAR data have also been conducted in recent years. Yang et al. and Bai et al. estimated SM in alpine grassland environments in Nagqu and Magu, respectively, both with favorable results [35,36]. In the Beiluhe, Zhang et al. used estimated SM in alpine meadows and alpine deserts and improved the accuracy of estimating SM by the WCM and CD algorithms [45]. However, these SM retrieving studies on the QTP were conducted outside or in the margin/border of the permafrost area or within a very small permafrost region. The study area was very small, i.e., covering only a few square kilometers, and the effectiveness and accuracy of the retrieving algorithms or models were not tested in other areas. To our knowledge, no SM retrieving study has ye<sup>t</sup> been conducted on the large-scale hinterland permafrost regions.

The freeze–thaw cycle of active layer soils and the water barrier of the frozen layer in the permafrost area play an active role in determining vegetation growth and SM retention [46,47]. It implies that in areas with high SM, the vegetation cover has significant interference with the radar signal. The vegetation canopy complicates the extraction of underlying soil water, as the canopy contains water and can also block or scatter radar signals [48]. Therefore, in the retrieval process, vegetation is another important factor that affects the radar signal in addition to SM. In the many ecological, hydrological, and agricultural studies, the vegetation canopy is usually expressed by the vegetation indices, such as Leaf Area Index (LAI), Normalized Difference Vegetation Index (NDVI), and the Enhanced Vegetation Index (EVI), biomass, vegetation height, etc. [35,45,49–52]. Several studies have shown that NDVI is easier to derive and has fewer errors than other vegetation indices and is widely used in SM retrieval studies [50,53]. In addition, in partially vegetated areas, Bao et al. found that Normalized Difference Moisture Index (NDMI) can also perform well in SM retrieval studies based on S1 data [54].

In the permafrost region on the QTP, the ground has a distinct freeze–thaw cycle process. The soil water is in a liquid state during summer and in a state of combination of ice and unfrozen water in other seasons [9]. The real implication of SM values obtained by in situ monitoring and sampling drying measurements could be different. In situ monitoring measures the unfrozen water (liquid water) content by monitoring the dielectric constant in the soil, such as the Hydra soil moisture sensor. The field-oven sampling acquires the total soil water content (unfrozen water and ice) by collecting in the field and then calculating the volumetric water content from the wet and dry weight of the soil. Meanwhile, *σ*◦ is sensitive to unfrozen soil water, and the frozen part is neglected in the retrieval process. The *σ*◦ could not represent the gross soil water content in all seasons except for the thawing season. Therefore, we need to be careful when choosing field "SM" data in developing and training SM retrieving algorithms [55].

In summary, a retrieval algorithm for SM is urgently needed to obtain SM spatial data which could promote hydrological, ecological, climatic, and engineering studies in the permafrost region of the QTP. In this study, the hinterland of the QTP was selected for SM retrieval, where a variety of surface types are included. The retrieval algorithm is trained and validated using multi-year in situ observations of different surface environments. We chose the months of July and August as the study period, which can reduce the errors caused by the freeze–thaw process of the soil. The SM retrieval during the thawing

season could represent the gross soil water content. In addition, the CD algorithm is a promising method for SM retrieval in the permafrost region where a priori knowledge is scarce [39,40]. The liquid soil water is very small in the coldest winter season, by reference to the CD algorithm, the effect of surface roughness is minimized. Then, the vegetation effect is further represented and reduced by vegetation indexes (NDVI, NDMI) from optical data. Finally, the objectives of this study are: (1) to develop an SM retrieval algorithm suitable for permafrost environments on the QTP using high spatial-resolution SAR data to obtain spatial data of SM for the thawing season; and (2) to explore the spatial–temporal distribution characteristics of SM on the large extent of permafrost region on the QTP.

### **2. Materials and Methods**

### *2.1. Study Area*

In this study, an area of 505 km × 246 km in the hinterland of QTP along the Qinghai– Tibet Highway was selected as the study area, as shown in Figure 1. This area covers most of the stations of the SM and temperature monitoring of the permafrost networks [11]. The study area includes typical permafrost regions and seasonally frozen ground regions with an altitude between 4189 m and 6402 m a.s.l. The average annual temperature in this region is between −5.8 ◦C and −2.4 ◦C, and the trend of temperature increase is consistent, with an average rate of change of about 0.05 ◦C/a. The annual precipitation is in the range of approximately 210–580 mm, with sizeable interannual variation [56,57]. The precipitation is mainly concentrated between May and September, and there is an apparent upward fluctuation in annual precipitation, with an average variable rate of 7.49 mm/a from 2004 to 2016 [58].

The vegetation types in the study area are classified as alpine swamp meadows, alpine meadows, alpine steppes, and alpine deserts. The alpine meadows cover the largest areas, followed by alpine steppes and alpine swamp meadows at the least [59]. The degradation of the permafrost has affected the ecological situation in the QTP. The vegetation ecosystem degradation is significant, mainly manifested as the degradation of alpine swamp meadow to the alpine meadow and alpine meadow to alpine steppe [59].

The frozen ground undergoes seasonal freeze–thaw cycles [8]. On the QTP, the thawing process begins in mid-to-late May and lasts until late September to early November each year [58,60,61]. The effect of water transport during soil freezing and thawing on SM distribution is very significant. The measured data in the in situ show that the SM in the thawing season varies roughly between 0.1 m3/m3 and 0.5 m3/m3. When the soil is frozen in winter, the unfrozen water content is low.

**Figure 1.** The overview of the study area. The base map is the map of the permafrost [10], glaciers [62], lakes [63], and topographic map from SRTM data [64].
