1. Introduction
Surface emissivity represents the surface thermal radiation capacity and is an important physical quantity in understanding the surface energy budget and surface radiation process. The snow-free land surface emissivity can not only separate the surface information from satellite observations but also improve the inversion accuracy of atmospheric and land surface parameters (such as water vapor [
1,
2,
3,
4], precipitation [
5,
6,
7,
8], soil water content [
9,
10,
11,
12], vegetation moisture content [
13], land surface temperature [
14], snow and ice cover [
15], etc.). In studying land surface emissivity, microwave (MW) and infrared (IR) regions of the electromagnetic spectrum are usually used. However, the MW region has become important in surface remote sensing owing to its long wavelength that can penetrate clouds and is affected less by the atmosphere. MW surface emissivity has a significant role in the global weather forecasting system and is needed to assimilate the MW radiation data of various satellites [
16,
17].
Recently, numerous studies have focused on exploring the use of MW land emissivity calculation methods from satellite measurements and physical models. Inversion methods are based on satellite data. Those studies employ methods that are empirical statistical [
18], semi-empirical model [
19], radiative transfer [
20,
21,
22,
23], exponential analysis [
24], neural networks [
25], one-dimensional variational [
26,
27,
28], parameterized [
29,
30] and so on. Comparisons of emissivity inversions by the various methods (physical modeling, statistical modeling, and hybrid of physical and statistical modeling) indicate that results are highly dependent upon the method used [
31,
32]. Those methods can directly calculate the land surface emissivity at large spatial and temporal scales. However, due to regional and seasonal influences as well as those of the input parameters, it is difficult to guarantee the results’ accuracy and stability.
For example, Jones and Vonder Haar [
33] retrieved MW land emissivity in the central United States. Prigent et al. [
20] calculated the Special Sensor Microwave Imager (SSM/I) emissivity of the land surface in most parts of Europe and analyzed its variation characteristics under different surface conditions. More recently, Zhang et al. [
34] analyzed the characteristics of MW emissivity for a variety of surfaces and the seasonal variation of the different surface types under different wavebands and polarization conditions. In another study, Prakash et al. [
35] found that the surface emissivity at horizontal polarization generally increases with the increase in frequency under clear-sky conditions. In particular, the surface emissivity of vegetation-covered areas showed obvious seasonal variations. More recently, Wu et al. [
36] used the radiative transfer equation to invert the surface emissivity of the Qinghai-Tibet Plateau.
The radiative transfer model (RTM) is basically used as a forward operator to simulate observation equivalents and to calculate the difference between these and the observations [
37]. Observations of brightness temperature (TB) that deviate too far from the simulated values are usually excluded from the assimilation. Therefore, the accuracy of forwarding TB simulation plays an important role in data utilization and subsequent assimilation effects. Numerous studies have also focused on physical surface emissivity models. Development of such models includes the simulation of bare land surfaces [
38,
39], vegetation-covered land surfaces [
40,
41,
42], snow cover [
43,
44], and complex land surface types [
45]. These physical models of MW emissivity for different surface types all possess certain physical meanings. However, some of the surface parameters in those models are often difficult to obtain, and inevitably, the high uncertainty of those parameters ultimately affects the accuracy of retrievals.
The uncertainty associated with land surface emissivity greatly reduces the use of spaceborne MW radiometer data. Thus, it is a requirement to improve existing MW land surface emissivity models further and develop new algorithms. At present, the largest spatial variability of surface emissivity occurs in bare rock or soil regions. However, these regions are simply divided into a wasteland, desert, and bare soil, affecting studies on the spectral response characteristics of the soil structure or rock minerals to the satellite-borne sensor. The deviation caused by the uncertainty of model parameters greatly affects the assimilation of MW observation data in desert areas. The latter seriously hinder any effective use of these data for this type of land surface. Therefore, it should improve MW observation data assimilation in the desert (or bare soil) areas and improve the utilization of MW data. To do this, it is necessary to provide further systematic and quantitative analysis of existing surface parameters and the physical properties of the bare soil emissivity model itself.
In view of the above, the present study analyzes the spectral characteristics of surface emissivity according to soil classification over the desert. A further objective has been to analyze the influence of mineral materials and soil texture information (e.g., soil component content, distribution of particle size) on the Microwave Land Emissivity Model (MLEM) simulation results. In this study, Microwave Radiation Imager (MWRI) data is used, to our knowledge for the first time, to invert the surface emissivity spectrum over the Taklimakan Desert. According to the Food and Agriculture Organization (FAO) [
46] and State Soil Geographic (STATSGO) dataset definitions of soil type, the desert surface emissivity spectra calculated by inversion are classified and analyzed. Following this, an MLEM [
45] is employed to calculate the emissivity spectra of the four most widely used materials in the desert area—namely, quartz, sandstone, granite, and limestone [
47]—under different particle sizes, soil components, and soil water content. The results are discussed, analyzing the main factors contributing to the error in calculating the MW emissivity of the desert area from the examined herein model. Also, possible ways to improve models’ predictions are proposed. As the MLEM is an important part of the global meteorological data assimilation and weather forecast system, this study’s findings can be valuable for improving the utilization of satellite data in desert areas and also expanding the models in numerical weather forecasting.
3. Methods
3.1. Inversion Method for MW Land Emissivity
There are some MW bands with weak atmospheric absorption, called atmospheric MW window regions. The commonly used atmospheric MW window areas include frequencies of 10 GHz, 19.4 GHz, 31~37 GHz, and 90 GHz. Because the observed values of the channel in the MW window region are least affected by atmospheric absorption, scattering, and emission, the inversion of surface emissivity is more accurate than results from non-window channels. Assuming that the surface is flat and the atmosphere is a plane parallel model and non-scattering medium, the Rayleigh-Jeans approximation can be employed for calculating the surface emissivity by using the satellite TB through the radiative transfer equation [
21,
45]:
where
Tb is the satellite TB and
TS is the temperature of the land surface, which can be obtained from NCEP/FNL datasets.
Tu is the satellite’s upward radiation and can be written as [
49]:
Td is the atmosphere downward radiation and can be written as [
49]:
Γ is the transmittance of the atmosphere and can be written as [
48]:
In Equations (2) and (3),
B is the brightness temperature of each layer with optical depth
τ,
τ0 is the optical depth at the top,
τs is the optical depth at the bottom of the layer,
μ = cos(θ), where θ is the satellite view angle.
Tu,
Td, and Γ are calculated from the NCEP-FNL temperature and water vapor profiles as inputs to the MW absorption model. Equation (1) indicates that the deviation of
Tb,
TS, and the atmospheric temperature/humidity profile may affect the accuracy of the inversion [
50]. It is known from the literature that an error of 1% in
Tb may lead to an emissivity error of more than 1%, depending upon the upwelling and downwelling radiative components [
49]. Furthermore, a 1% error in
TS would result in a 1% error in land emissivity. Moreover, the error in
TS is the main error source for emissivity at frequencies less than 19 GHz [
50]. In addition, the error in atmospheric profiles has a greater impact at higher frequencies, where emission-based radiative transfer could be more problematic.
3.2. Microwave Land Surface Emissivity Model
Weng et al. [
45] developed an MLEM using a two-stream approximation. In the model, the surface of Earth is depicted as a multi-layered medium on an irregular surface. The top and bottom layers are taken to be interstitially homogeneous, while the intermediate layers are interstitially heterogeneous and contain scatterers such as snow, sand, and vegetation. At present, MLEM already contains most of the important radiative transfer processes on the land surface so that it can be applied to most of the land surface conditions around the world.
Although this model can simulate the emissivity well for a variety of surface conditions, the level of deviation in the calculation for desert areas is quite large. One of the reasons for this is the uncertainty attached to some of the important parameters in the model. For example, the geometric scale (the shape and size of sand grains and snow crystals, the height of vegetation canopy, blade shape and inclination, the shape of branches and vegetation fraction) of the scatterers (e.g., snow, sand, and vegetation) is very difficult to obtain from conventional observation data. For those parameters, therefore, the model generally assumes them to be constants. For example, the contents of clay and sand of different types of soil are set to a fixed value in the model. Actually, for each type of soil, there is a range of sand content or clay. However, it is difficult to obtain accurate sand content of different types of soil. In addition, the particle radius for different types of soil was set as a fixed value in the model. Nevertheless, there is a radius range of sand or clay particles. Therefore, the same type of soil may have a different distribution of particle size. Evidently, such set values for those model parameters do not apply to all types of soil and thus become the main error source in the current model. To reduce the impact of parameter uncertainty on the model and use the model more effectively to analyze the influence of soil texture on surface MW emissivity in more detail, herein information on soil texture, particle sizes, and sand/clay content are obtained from the FAO/STATSGO soil texture database. Then, it is used the U.S. Department of Agriculture’s (USDA) classification scheme to define the ratios of clay, silt and sand to different soil types and adjust the parameter values corresponding to different soil types in the model.
3.3. Verification of MLEM Simulation Results
A key study objective has been to analyze the correlation between emissivity and different soil textures and further verify that the addition of soil texture information (such as sand fraction, clay fraction, distribution of particle size, and so on) improves the accuracy of the model simulation results. To do this, the emissivity retrieved from satellite data is used as the “true value” to calculate the deviation between the retrieved emissivity value and the simulated value. The MLEM simulates the emissivity of the Taklimakan Desert combined with GDAS data. The simulation results are compared with the emissivity retrieved by FY-3B/MWRI measurements. The roughness of the satellite remote sensing is actually related to the spatial resolution of the remote sensor. According to different frequency ranges, the model gives the dependence functions of horizontal and vertical polarization emissivity on surface roughness. The output of GDAS provides global surface characteristic parameters (such as surface temperature, soil temperature, surface vegetation type, canopy moisture content, etc.) for the model. If the emissivity deviation after considering the difference in soil texture is smaller than that of the original model, it suggests that the refinement of soil texture information in the model can improve the model’s accuracy.
6. Conclusions
In this study, MWRI TB data is used, to our knowledge for the first time, to calculate the surface emissivity under clear-sky conditions in the Taklimakan Desert region. Combined with soil texture classification information from the FAO and STATSGO, the variations of MW emissivity for different soil types in the Taklimakan Desert are analyzed. In addition, using the MLEM, the emissivity spectra of the four most abundant materials in the desert area—namely, quartz, granite, sandstone, and limestone—under different particle sizes, soil components, and soil moisture content are calculated. Then, on this basis, the soil texture error sources used in the current MLEM’s calculations in desert areas are discussed. Moreover, according to the soil texture information, the feasibility of reducing the MLEM simulation error in the desert area by refining the soil classification characteristic parameters in the model has also been verified. The key study findings are summarized below:
The MW emissivity in desert areas is highly correlated with the soil type and the seasonal variation of land surface emissivity for different soil types differs considerably. The seasonal variation of surface MW emissivity of clay-rich soil is more obvious than that of sand-rich soil.
Soil moisture is affected by precipitation to some extent but is also restricted by soil type. This is because the water content of different types of soil is different on the whole due to the difference in water storage capacity.
The surface emissivity changes considerably with a difference in the soil distribution of particle size. For the same mineral, the horizontal polarization emissivity generally decreases with the increase in soil particle radius. Furthermore, the emissivity of soil composed of small-size particles has marked seasonal characteristics, and the emissivity of the horizontal polarization shows stronger seasonal variation than that of the vertical polarization.
In the desert surface layer, where the soil is mainly sandy in type, the surface emissivity is affected by the depth of the desert to some extent. Because soil moisture in desert areas is very low throughout the year, the penetration depth of soil is an important factor affecting the surface emissivity.
The fact that surface emissivity is dependent on soil texture requires the theoretical model to consider the influence of soil texture in its practical application. The increase in soil texture information, including finer details regarding the soil composition content and distribution of particle size, improved MLEM’s simulation in the desert region—especially for desert soil containing a large number of small particles. The simulation error of the model after adjusting the parameters is considerably reduced.
It is important to acknowledge that in this study, data from one calendar year are used to study the MW emissivity of the land surface in the Taklimakan Desert. Therefore, further analyses over longer time ranges should be carried out as it would allow for providing more comprehensive emissivity variation characteristics in this area. At the same time, the specific reasons behind the spatial and temporal variations of the emissivity and the differences between ascending and descending orbits need to be analyzed and verified by combining more observational and experimental data. It is possible to improve the existing model by analyzing the radiative transfer process by stratifying the near-surface soil. From the preliminary results of this study, it is clearly evidenced that the MLEM will be improved since it can be seen that soil composition (e.g., sand fraction, clay fraction), mineral dielectric constant (e.g., quartz, sandstone, granite, limestone, etc.) and soil particle size all affect the simulation results of the emissivity. As the MLEM is an important part of the global meteorological data assimilation and weather forecast system, the model’s prediction improvement will enhance the use of satellite data in desert areas and will improve the accuracy of the numerical weather forecast. These aspects remain to be seen.