2.5.1. Theoretical Framework for Spectral LSE and LST Inversion
The calculation scheme used for the verification of temperature and emissivity inversion for different land surface types is shown in
Figure 6.
The overall calculation process, as shown in
Figure 6, is as follows: First, the measurement data undergoes noise removal using a Gaussian filter, and the atmospheric transmission characteristics parameters are calculated using Modtran4.0 to perform atmospheric correction on the surface medium-wave high-spectral imaging data. Then, Modtran4.0 is utilized to calculate the solar irradiance and atmospheric background radiation. The surface bidirectional reflectance distribution function (BRDF) model is applied to mitigate the scattering effects of sunlight and atmospheric background radiation on the surface.
Considering the rectangular area of the experimental observations, it can be approximated as a point in relation to the 5.56 km spatial resolution of MODIS imagery. In order to address the spatial resolution disparity between ground and satellite scales, this experiment selects sampling sites with high homogeneity and good spatial representativeness. The point-to-area upscaling method is employed [
31], where the values of ground sampling points are directly averaged in order to match the spatial resolution of the ground observation data.
Subsequently, a comparison and a validation are conducted by comparing the redistributed satellite data with the ground observation data. Finally, the processed data is used for spectral land surface emissivity (LSE) and land surface temperature (LST) inversion, and the results are compared and validated using MODIS products.
- (1)
Noise Estimation and Removal
- (a)
Sensor Noise Estimation
The noise level performance was evaluated in the laboratory using the noise equivalent spectral radiance (NESR) metric to assess the noise level of the Hyper-Cam MW sensor. Based on the imaging principle of the Hyper-Cam MW sensor with a specific optical system, the NESR is related to the spectral resolution and integration time of the Hyper-Cam MW sensor. In the noise level estimation, NESR is calculated as the pixel-wise standard deviation of the radiance from multiple consecutive data cubes of the same target [
24]. The noise calculation results are shown in
Figure 7.
Figure 7 showed that after performing the average spectral NESR calculation, within the range of 3 to 3.5 μm, the NESR increases with increasing wavelength, and there is a relatively high level in this range. Near 4.2 μm, there is a peak, which is caused by a significant increase in path radiance due to the presence of a carbon dioxide atmospheric absorption band in the vicinity of that wavelength.
- (b)
Sensor Noise Removal Based on Gaussian Filtering
Both the entire mid-wave infrared hyperspectral image and the noise within random windows approximate a Gaussian distribution. Therefore, Gaussian convolution is applied as a spatial denoising method in order to improve the noise quality in each spectral band. This approach takes into account that the noise type in small windows also approximates a Gaussian distribution. It calculates a weighted template using the Gaussian kernel model and then convolves the weighted template with the entire moving window of the hyperspectral image. The window size of the Gaussian template is 3 × 3 pixels. Gaussian filtering for spatial denoising can effectively remove noise from the experimental image while preserving the atmospheric features present in the captured data [
24]. The denoising results are shown in
Figure 8 and
Figure 9.
Figure 8 and
Figure 9 showed that after applying Gaussian spatial convolution filtering, the regions with correlated noise in the original measured image are significantly removed, and the NESR shows a noticeable decrease. This process enhances the image quality of the hyperspectral data, resulting in improved clarity and reduced noise interference.
- (2)
Atmospheric correction:
During the experiment, the radiation detected by the sensor includes three components: atmospheric path radiance, scattering of the target’s own radiation by sunlight and atmospheric background radiation, and the target’s own thermal radiation. The spectral radiance received by the detector is expressed in Equation (1).
In order to remove the influence of atmospheric transmission, atmospheric correction is applied to the measured data so as to eliminate the effects of atmospheric path radiance and transmission attenuation, thereby obtaining the apparent radiance of the target. The processing method is shown in Equation (2).
According to Equation (2), the atmospherically corrected data include the target’s thermal radiation and the scattering of the target’s radiation by sunlight and atmospheric background radiation. In order to accurately invert the emissivity, it is necessary to remove the scattering of the target by sunlight and atmospheric background radiation, thereby obtaining the target’s own thermal radiation. The calculation expression for the target’s self-radiation characteristics is shown in Equation (3).
In Equation (3), represents the measured surface spectral radiance, represents the solar zenith angle under actual terrain conditions, represents the surface anisotropic reflectance, represents the surface hemispherical-directional reflectance, represents the solar irradiance, represents the sky background radiance, represents the transmittance of the detection path, and represents the radiance of the detection path.
All atmospheric transmission parameters, such as atmospheric path radiance, atmospheric transmittance, solar radiation irradiance, and sky background radiance, are calculated using Modtran4.0. The surface anisotropic reflectance and surface hemispherical–directional reflectance can be obtained through inversion using MOD43C1 products.
- (3)
Spectral Land Surface Emissivity (LSE) Inversion:
The spectral LSE is inverted using the Planck blackbody radiation formula. The inversion equation for spectral LSE is shown in Equation (4).
In Equation (4), represent the thermal radiation spectrum of the surface background, represent the temperature (in Kelvin), and represent the wavelength (in µm), respectively.
- (4)
Land Surface Temperature (LST) Inversion:
The equation for LST inversion is shown in Equation (5).
In Equation (5), represent the surface emissivity, represent the thermal radiation spectrum of the surface background (in W/(m2·sr)), represent the temperature (in Kelvin), and represent the wavelength (in µm), respectively.
2.5.2. Acquisition of Surface Background BRDF
In order to remove the scattering component of the target by sunlight and atmospheric background radiation, the anisotropic scattering of the surface needs to be considered. A linear kernel-driven BRDF model is used to solve this problem. For the two natural land surface types (grassland and desert), BRDF parameter data of the surface can be obtained from the MCD43C1 product. The MCD43C1 product provides global surface reflectance, bidirectional reflectance distribution function (BRDF) parameters. These parameters are derived from data collected by the Terra and Aqua satellites of the MODIS (moderate resolution imaging spectroradiometer) system. The spatial resolution of each pixel is 0.05 degrees. The MCD43C1 data product is generated through processing and analysis of the radiance measurements in different spectral bands. This product provides the linear weighting coefficients of the RossThick-LiSpare-Reciprocal BRDF model for bands 3 to 5 μm [
32]. The surface reflectance is a linear combination of isotropic scattering, volumetric scattering, and geometric optical surface scattering [
33,
34], as shown in Equation (6). Using this equation, the bidirectional reflectance at any incident and viewing angles can be obtained.
In Equation (6), represent the incident zenith angle, represent the viewing zenith angle, and represent the relative azimuth angle between the viewing direction and the radiation incident direction, respectively. Additionally, ,, represent the proportion coefficients of isotropic scattering kernel, volumetric scattering kernel, and surface scattering kernel in the linear combination, which can be obtained from the MCD43C1 product.
In MODIS products, the isotropic scattering kernel is defined as
. For the volume scattering kernel, the volumetric scattering kernel model proposed by Roujean is used [
33]. This model mainly describes densely vegetated layers with similar distribution of leaves, implying that their reflectance can be considered somewhat equivalent. The expression for the volume scattering kernel is as follows [
35]:
represents the phase angle, which is related to the solar zenith angle, the viewing zenith angle, and their relative azimuth angle:
For the geometric optical surface scattering kernel
, considering the mutual occlusion between vegetation canopies, the modified LiSparse [
34] model proposed by Lucht is used. The specific expression is as follows:
represents the overlap region between the observation and the solar shadow, and the specific formula is:
The expressions for other parameters are as follows:
In MODIS products, the parameter
is defined as a dimensionless parameter representing the relative height and shape of the canopy top. Combining Equations (6) to (14), the linear weighting coefficients of the BRDF for different land surface backgrounds in 2020, obtained from the MCD43C1 product, are shown in
Figure 10.
Figure 10 showed that the isotropic scattering kernel weight (
) contributes the most to both land surface types, and it exhibits significant fluctuations throughout the four seasons. This can be attributed to the relatively uniform biological and physical structure of the land surface vegetation, allowing the weight coefficient to effectively capture the temporal variations of the actual land surface biophysical properties. The contribution of the volume scattering kernel (
) shows slightly smaller fluctuations compared to the isotropic scattering contribution (
), which can be attributed to the inherent differences in biological properties and structural characteristics of barren land and grassland. The Fvol values remain relatively low throughout the year. The contribution of the geometric scattering kernel (
) is the lowest for both land surface types, representing the correction for mutual occlusion. It is noteworthy that the
values for grassland and barren land tend to approach zero, validating this conclusion.
By utilizing the geometric positions of the sun and the detector, as shown in
Table 5, and combining them with Equation (6), the anisotropic reflectance of the land surface can be calculated. The calculated BRDF values for the grassland and the Gobi Desert are 0.043 and 0.075, respectively.
2.5.3. Typical Surface Emissivity and MODIS Reflectance Data Product Retrieval
- (1)
Surface Emissivity Retrieval
Emissivity data for the desert and grassland in the northwest plateau region of China in 2020 were obtained from the MOD11 product. These data are used for comparative validation of the spectrally emissivity inversion obtained from the mid-wave hyperspectral imaging measurements. The MODIS product provides three narrow-band reflectance data in the mid-wave infrared region, namely Band20 (3.66~3.840 μm), Band22 (3.929~3.989 μm), and Band23 (4.010~4.080 μm). The temporal variation of emissivity for grassland and desert in these three bands in 2020 is shown in
Figure 11.
According to the analysis from
Figure 11, the emissivity of grassland surface varies between 0.934 and 0.985 throughout the year, showing consistency across the three mid-infrared bands. In March, it reaches the lowest value of 0.934, while it reaches the maximum value in July. From July to December, the emissivity values gradually decrease. In August, he mid-infrared emissivity values for grassland are 0.968, 0.96, and 0.956. On the other hand, the emissivity of barren land surface fluctuates between 0.806 and 0.970, showing larger variations throughout the year. In September, the mid-infrared emissivity values for barren land are 0.9545, 0.93, and 0.886.
- (2)
Retrieval of Hemispherical-Directional Reflectance for Natural Surfaces
According to the directional Kirchhoff law [
36], for opaque objects, the relationship between the directional emissivity and the hemispherical–directional emissivity can be expressed as:
By combining Equation (6), the mid-wave infrared hemispherical-directional emissivity can be obtained as:
From Equation (8), it can be observed that, based on the weights
, the hemispherical-directional emissivity of the surface can be obtained at different angles. The variation trends of the hemispherical–directional emissivity for grassland and desert in the infrared wavelength range during different seasons are shown in
Figure 12, considering the weighting coefficients provided in
Figure 10.
Figure 12 showed that the mid-wave hemispherical emissivity of grassland and barren land surfaces shows a similar trend. It remains relatively stable for observation zenith angles between 0° and 40° and starts to decrease from 40° to 80°. From a temporal perspective, there is a distinct difference in the grassland environment during winter compared to other seasons. The hemispherical emissivity of grassland is the lowest in winter, ranging from 0.48 to 0.53, while in the other three seasons, it ranges from 0.72 to 0.78. On the other hand, the hemispherical emissivity of barren land fluctuates between 0.5 and 0.7, with the highest values being observed in summer, followed by autumn, winter, and spring. This phenomenon is mainly related to vegetation characteristics and properties.