-
In the study, the FY-3B MWRI (Table 1) Level 1 recalibrated swath brightness temperatures (ascending orbits) were obtained from the National Satellite Meteorological Centre for 7 July 2015 through 30 June 2019. The spatial resolution varies between frequencies, and the instantaneous field of view (IFOV) sizes range from 51 km × 85 km at 10.65 GHz to 9 km × 15 km at 89.0 GHz. The historic MWRI raw data were generated into the fundamental climate data record (FCDR) with the latest preprocessing algorithms; the dataset underwent careful corrections in terms of geolocation errors (Tang et al., 2016; Li et al., 2020b; Liu et al., 2021) and issues of intrusion of the hot reflector, which was not accurately estimated in the prelaunch calibration process and resulted in a slight difference between the ascending and descending orbits (Xie et al., 2019). In addition, the MWRI data went through series of space-based cross-calibrations with SSMIS and TMI (Zhang et al., 2019), and, as is reported, most of the biases are within 1–2 K. The cross-calibrations of MWRI with AMSR-E and AMSR2 were conducted using the double difference method on the overlapped records (Du et al., 2014; Wu et al., 2020).
Frequencies (GHz) 6.925 10.65 18.7 23.8 36.5 89.0 MWRI sensitivity (K) − 0.5 0.5 0.5 0.5 0.8 AMSR-E sensitivity (K) 0.3 0.6 0.6 0.6 0.6 1.1 Table 1. The sensitivities of FY-3B MWRI and AMSR-E.
Himawari-8 AHI identifies sky conditions (clear or cloudy) by utilizing a cloud screening algorithm (Ishida and Nakajima, 2009) on the AHI L1B data and determines the cloud phase according to the brightness temperature (Tb) at different IR bands and their differences (Baum et al., 2000). In the AHI L2CLP cloud product, QA flags use two-bit values to provide information on the cloud retrieval quality confidence, cloud mask confidence level, cloud retrieval phase flag, and so on. Cloud retrieval phase flags were identified as “clear”, “liquid”, “mixed or uncertain”, or “ice”. The measurements covered the majority of China with a nominal coverage of 60°S–60°N, 80°E–160°W.
As shown in Fig. 1, we used a Microwave Radiative Transfer (MWRT) model (Liu, 1998) to simulate the upwelling microwave brightness temperature at the top of atmosphere (TOA) measured by satellite. The model is a four-stream radiative transfer model calculating Tbs for plane-parallel atmosphere using surface-, atmospheric-, and precipitation-related variables. It has been widely used in numerous studies (Liu, 2004; Noh et al., 2006, 2009; Okamoto et al., 2008; Aonashi et al., 2009a, b; Fujii et al., 2009; Shige et al., 2009, 2013; Wang et al., 2009, 2019a; Taniguchi et al., 2013; Kim et al., 2015; Jeoung et al., 2020; Kubota et al., 2020; Wu et al., 2020) for over 20 years and is actively being updated with new features. It is also used to calculate microwave single-scattering properties for non-spherical ice particles (Liu, 2004, 2008) as well as to retrieve snowfall (Noh et al., 2006, 2009; Liu, 2020). In the model, the contributions from land surface emission, the attenuation from atmospheric oxygen and water vapor, and the emission and scattering from cloud droplets were all included. Those fundamental inputs were derived from multiple data sources, including microwave measurements from the FY-3B MWRI, the observations and retrievals from visible and infrared sensors on the geostationary Himawari-8 satellite AHI, and the land surface temperature, the profiles of atmospheric temperature, water vapor, and geopotential height from ERA-Interim atmospheric reanalysis datasets (Berrisford et al., 2011). All pixels with Himawari-8-retrieved cloud water path over 300 g m–2 were assumed raining and were excluded from the MLSE retrieval process.
In this study, we carefully considered the cloud contributions to TOA Tbs by using fused Himawari-8-retrieved cloud properties. Analogous to the work in Han et al. (2016), we collocated the high-resolution AHI cloud observations to the large MWRI footprint with the method of arithmetic mean. When collocating AHI cloud properties and MWRI observations, differences in footprint sizes at different MWRI channels were considered so that the high-resolution AHI clouds were averaged inside the corresponding footprint of certain MWRI channels individually. When fusing the ERA-interim atmospheric profiles and land skin temperature with MWRI, two 6-h ERA estimates that temporally bracket the MWRI overpassing time and spatially encompass from the nearest grid to the center of the MWRI footprint are collocated. Cloud water path (CWP) was derived from COT and CER with
$ \mathrm{C}\mathrm{W}\mathrm{P}=2/3\mathrm{C}\mathrm{O}\mathrm{T}{\rho }_{\mathrm{w}}\mathrm{C}\mathrm{E}\mathrm{R} $ for liquid phase and$ \mathrm{C}\mathrm{W}\mathrm{P}={\mathrm{C}\mathrm{O}\mathrm{T}}^{1/0.84}/0.065 $ for ice phase, where$ {\rho }_{\mathrm{w}} $ is the density of pure water. The optical and infrared observations of Himawari-8 did not provide vertical information on cloud properties. Therefore, we assumed a single-layer cloud located at the ERA-interim atmospheric profiles layer determined by CTT. Then, the cloud water content (CWC) of different cloud phases was calculated by dividing the derived CWP with the thickness of the determined cloud layer. In Liu’s (1998) MWRT model, cloud water and cloud ice were modeled as spherical particles; both liquid and ice phase droplets were assumed following the general three-parameter gamma particle size distribution:$ N\left(D\right)={N}_{0}{D}^{\mu }{e}^{-\lambda D} $ . For liquid phase, λ was set to$ 3\times {10}^{5} $ and μ = 2; for ice phase, λ was set to$ 2.55\times {10}^{3} $ and μ = 0 according to Heymsfield et al. (2002). The$ {N}_{0} $ was derived with the input CWC and specified λ and μ. The mixed-phase cloud water was treated as half liquid and half ice phases, and the two different halves were processed individually. The dielectric constants of liquid water and ice water were based on Lieb et al. (1991) and Mätzler (2006).Based on the above cloud configuration, the absorption and scattering coefficients of cloud particles were calculated using Mie theory in the microwave radiative transfer model. Therefore, for given TOA Tbs over a cloudy pixel, the contribution from clouds could be quantified. After further calculating the contributions from water vapor and oxygen, the upwelling Tbs at land surface could be derived. Then, the retrieval algorithm iteratively adjusted the value of MLSEs until the simulated TOA Tb matched the real satellite measurements.
As shown in Fig. 1, there are three main steps throughout the retrieval algorithm. First, multisource datasets were collocated on a spatiotemporal basis and by screening rainy/water/ocean pixels. A threshold of 0°C was set for the surface skin temperature to roughly screen out surface snow/frozen pixels. Second, AHI cloud observations in the MWRI footprints were converted to associated parameters required in the MWRT model. Thirdly, multichannel MLSEs were retrieved using the iterative process described in Fig. 1.
This retrieval process is independent for each FY-3B MWRI sample and channel. The final product covers the majority of China and parts of neighboring countries. Temporally, it spans from 1 July 2015 to 30 June 2019.
For comparing purposes, we compared our MLSE retrievals with the previous MLSE products derived from Aqua AMSR-E observations to assess the product under cloud-free conditions. Norouzi et al. (2011) built an MLSE product with frequencies ranging from 6.9 GHz to 89 GHz (Norouzi_MLSE) under cloud-free conditions. Moncet et al. (2011a) derived an emissivity product dataset (Moncet_MLSE) with frequencies above 10.65 GHz. The primary input data sources for retrieving MLSE are listed in Table 2. It should be mentioned that the two AMSR-E MLSE products represent the land surface properties during 2002 to 2010, while our MLSE is for 2015–19. The land use and land changes during those years will certainly introduce differences into the MLSEs.
Product MWRI_MLSE
(clear & cloudy sky)Norouzi_MLSE
(clear sky only)Moncet_MLSE
(clear sky only)Duration 20150701–20190630 20020701–20080630 20030101–20031231 Tb FY-3B MWRI AMSR-E L2A AMSR-E L2A Atmospheric profiles ECMWF/ERA-Interim ISCCP/TOVS NCEP/GDAS LST ECMWF/ERA-Interim ISCCP-DX MODIS LST Cloud flag Himawari-8/AHI L2CLP ISCCP-DX MYD06_L2 Cloud properties Himawari-8/AHI L2CLP − − Table 2. Data sources of the three MLSE products discussed in this study.
In addition, the MODIS/Aqua Vegetation Indices Monthly L3 Global 0.05Deg CMG (MYD13C2) dataset (Didan, 2015) was adopted to provide normalized difference vegetation index (NDVI) information on a monthly basis. In our analysis of the correlation between monthly MLSE and surface precipitation, daily rainfall estimation was taken from 1°
$ \times $ 1° Global Precipitation Climatology Project (GPCP) Climate Data Record (CDR), Version 1.3 (Daily) (Adler et al., 2017). MODIS/Aqua Snow Cover Monthly L3 Global 0.05Deg CMG (MYD10CM), Version 6 (Hall and Riggs, 2016) was obtained to provide the surface snow cover extent. -
Figure 2 shows an example of instantaneous MLSE retrieval on 26 July 2018 in the study area. Because the horizontal polarized (H-pol) MLSE is more sensitive to the surface changes in soil and vegetation water content (Prigent et al., 2006), we will primarily focus on horizontal polarizations in the following sections.
Figure 2. FY-3B MWRI-based instantaneous retrievals of MLSE (horizontal polarizations) at (a) 10.65 GHz,(b) 18.7 GHz, and (c) 36.5 GHz along with (d) cloud water path (CWP) retrieved from quasi-simultaneous observations from Himawari-8 AHI on 26 July 2018.
Of the selected study area, Fengyun-3B MWRI observed 69.7% of the land area, of which 66.3% was covered by clouds. In the MWRI-observed area, we successfully retrieved MLSEs in 86.7% area under both clear sky conditions and cloudy sky conditions. If we exclude all cloudy pixels, as previous algorithms did, we could only retrieve MLSE for 20.4% of the observed area. It should be noted that the radiative effect from clouds is not linearly proportional to the cloud fraction. And for a given cloud fraction, different types of clouds (cirrus, deep convective clouds, etc.) can make very different impacts on the MW radiative transfer.
By comparing the spatial distribution of retrieved MLSE and cloud water path (CWP, Fig. 2e), we can see a continuous and smooth transition of MLSE values from non-raining cloud-covered areas to cloud-free areas, indicating that the contribution of clouds is successfully removed from the retrieval algorithm.
From the probability distribution function (PDF) of retrieved MLSE (Fig. 3), we can see most of the values fall in the dynamic range of 0.7–1.0, except for a few samples at vertical polarizations of 18.7 GHz and 36.5 GHz, which result from random errors. It indicates most of our MLSE retrievals are at least reasonable. Generally, the PDF of MLSEs at lower frequencies are more at the low end than those at higher frequencies. This is consistent with our understanding about the frequency dependence of MLSE.
-
The spatial distributions of FY3B-MWRI MLSE at three frequencies (10.65 GHz, 18.7 GHz, and 36.5 GHz, horizontal polarization) for multiple-year-mean Summer (JJA) and Winter (DJF) are shown in Figs. 4 and 5. The associated results from Aqua AMSR-E Norouzi_MLSE (Norouzi et al., 2011) and Moncet_MLSE (Moncet et al., 2011a) are also shown. It should be noted that FY-3B MWRI and Aqua AMSR-E operated in different years. Due to the limitation of Himawari-8’s coverage, the FY3B-MWRI MLSE to the west of 80oE was not retrieved.
Figure 4. Spatial distributions of MLSEs derived from (a, d, g) FY-3B MWRI_MLSE (2015–19 JJA); (b, e, h) Aqua AMSR-E Norouzi_MLSE (2003–08, JJA), and (c, f, i) Aqua AMSR-E Moncet_MLSE (2003 JJA).
In summer (Fig. 4), our results show that the spatial distributions in China of MLSE for multiple channels are generally high in the southeast vegetated area and low in the northwest barren, or sparsely vegetated, area. The maximum values of MLSEs (particularly for 18.7 GHz and 36.5 GHz) are located in the southwest–northeast oriented belt area which is composed of the Qinling-Taihang Mountains and the eastern edge of the Qinghai-Tibet Plateau, which is also the dividing belt between high and low MLSE in China. Very low MLSEs (e.g., 0.8–0.86) are seen in Xinjiang (e.g., the Taklimakan Desert), western Inner Mongolia, and the Yangtze River delta. In southeastern China, there are still high spatial variations of MLSE with moderate values in vegetated areas and low values in river basins, coastal areas, and croplands, such as the Yangtze River basin, Sichuan basin, and northeast plain. In terms of frequency dependence, MLSEs generally increase with increasing frequency. However, in densely vegetated areas, such as forests in the southeast coastal and northeastern areas of China, MLSE at a lower frequency (e.g., 18.7 GHz) may be larger than that at a higher frequency (e.g., 36.5 GHz). This is consistent with the results in Min and Lin (2006b), Min et al. (2010), and Li and Min (2013).
In winter (Fig. 5), the spatial patterns of the three MLSE products are generally consistent. Previous studies have shown MLSE at low (high) frequencies over mountains and high-latitude areas increases (decreases) due to the snow cover effect (Prigent et al., 2006; Ferraro et al., 2013). The MLSE over the southern forest area also increases a little in winter. The seasonal variations of MLSE are further studied in the following section.
Relatively large discrepancies are found between the MLSE products in southwestern China, where the high values of this study’s MLSE are mainly located to the west and south of the Sichuan Basin; the high values of Norouzi_MLSE are more to the north and closer to Lanzhou; the Moncet_MLSE exhibits moderate values in southwestern China. These discrepancies of pattern and magnitude are mainly due to the use of different land surface temperature sources, which are significantly different in winter in this region. Overestimations are found in ISCCP skin temperatures (used by Norouzi et al., 2011) compared to those of ERA (used by this study) in these two regions. ISCCP daytime skin temperatures can be +5.0 K higher than MODIS in July, and there are areas with differences as large as 25 K (Moncet et al., 2011b). Although it is hard to tell which one is more biased, ERA represents the full range of weather conditions, whereas ISCCP and MODIS will miss cloudy samples. Additionally, the spread of snow cover in the high latitude regions in winter adds to this difference, since the snow is not completely removed from our retrieval. As for the high values in the Qinling-Taihang Mountains, as a common feature in all three products, the surface roughness and relief effect play main roles.
Despite a significant difference of observing years, the spatial patterns of FY3B-MWRI MLSE are highly consistent with Aqua AMSR-E Norouzi_MLSE and Moncet_MLSE. Detailed statistics of the correlation coefficient, root-mean-square error (RMSE), mean bias (MB), and mean absolute error (MAE) are shown in Table 3. As shown in Table 3, the spatial correlation coefficients between MWRI MLSE and the two AMSR-E MLSEs are 0.87–0.92 for horizontal polarizations and 0.67–0.83 for vertical polarizations in summer. The RMSEs are 0.016–0.022, and the MBs are –0.001–0.024, which are close to the differences between model-simulated MLSE and satellite retrievals reported by Prigent et al. (2015).
Frequency (GHz)/Polarization MWRI vs. Norouzi MWRI vs. Moncet Samples R RMSE MB MAE Samples R RMSE MB MAE 10.65H 15028 0.917 0.020 0.007 0.015 15348 0.904 0.022 0.015 0.021 18.7H 15028 0.920 0.019 0.002 0.013 15348 0.922 0.019 0.010 0.018 36.5H 15028 0.867 0.022 0.021 0.026 15348 0.898 0.020 0.020 0.023 10.65V 15043 0.748 0.016 0.005 0.012 15388 0.828 0.016 0.010 0.014 18.7V 15043 0.726 0.017 −0.001 0.012 15388 0.825 0.016 0.001 0.013 36.5V 15043 0.669 0.018 0.024 0.027 15388 0.808 0.017 0.016 0.019 Table 3. The statistical metrics of the intercomparison between MWRI MLSE and two AMSR-E-based MLSEs.
The high consistency between our MLSE and the two selected AMSR-E-based MLSEs indicates that the changes of land surface radiative properties during 2002 to 2019 are slow. It also demonstrates the measurements of Tb, including its calibration and validation by FY-3B MWRI, are reliable, and the retrieval algorithm developed in this study works well.
-
In a given region, the temporal variation of MLSE can reflect the changes of multiple land surface properties, including soil moisture, vegetation, precipitation, snow cover, etc. In order to study the temporal variations of MLSE in China with different types of land cover, we selected eight regions (Fig. 6, Table 4), including two mixed forests (A, B), two croplands (C, D), two grasslands (E, F), one island with mixed tropical forest, savannas (G), and one desert (H). Time series analyses (Figs. 7 and 8) of MLSE were performed in these regions on monthly basis along with NDVI, GPCP monthly precipitation, and MYD13C2 snow cover.
Figure 6. Selected regions for studying seasonal variations of MLSE in China. Overlapped is the IGBP land cover type in 2010 from MCD12C1 (Sulla-Menashe and Friedl, 2015).
Regions Vegetation zone Longitude (E)/
Latitude (N)Proportions of different land types (%) Barren Forests Crop Grass and
shrubsSavannas Urban and
build-upsWater A Subtropical evergreen broadleaf forest 106°–112°/22°–25° <0.01 33.1 16.9 0.04 49.1 0.61 <0.01 B Evergreen coniferous forest 119°–125°/50°–53° <0.01 93 4.9 1.56 0.45 0.06 <0.01 C Temperate deciduous broadleaf forest 114°–118°/35°–38° 0.02 0.8 97.8 0.23 <0.01 1.14 0.27 D Temperate grassland 123°–128°/47°–49° <0.01 1.0 97.1 1.22 <0.01 0.22 0.38 E Qinghai-Xizang Plateau vegetation 97°–102° /30°–34° <0.01 5.0 <0.01 95 <0.01 <0.01 <0.01 F Temperate grassland 107°–110°/37°–40° <0.01 <0.01 0.89 99.1 <0.01 <0.01 <0.01 G Tropical rain forest 108°–111°/18°–20° 0.08 28.4 38.6 <0.01 32.3 0.17 0.11 H Temperate desert 80°–84°/37°–40° 99.9 <0.01 0.02 0.03 <0.01 <0.01 <0.01 Table 4. Proportions of different land types in the eight selected regions.
Figure 7. Time series of monthly mean FY-3B MWRI_MLSEs (at 10.65H as blue lines, 18.7H as orange lines, and 36.5H as red lines), GPCP rainfall (sky blue bars), vegetation greenness (NDVI as green lines), and snow cover (purple bars) from July 2015 to June 2019 in selected regions (a) Region A (forest in southern China), (b) Region B (boreal forest), (c) Region C (croplands in the North China Plain), (d) Region D (boreal croplands), (e) Region E (grasslands over the Tibet Plateau), (f) Region F (grasslands over the Loess Plateau), (g) Region G (tropical forest on Hainan Island), and (h) Region H (desert in Xinjiang).
Figure 8. Probability Distribution Function (PDF, %) of MLSEs in winter (DJF) and summer (JJA) at three frequencies (at 10.65H in blue, 18.7H in orange, and 36.5H in red) in the eight selected regions.
In the southern forest Region A (Fig. 7a), the amplitude of the seasonal variation of MLSE is nearly the smallest among all the regions. Generally, the dependence of MLSE on vegetation (indicated by NDVI) and precipitation is not very clear in this region, although we can see the MLSE in winter is a little smaller than that in summer due to the growth of vegetation. In addition, the frequency dependence of MLSE in this region is very weak (Fig. 8a). This is unique because in most other regions, MLSE increases with frequency, except in the situation under snow cover (see further discussions).
In the boreal forest Region B (Fig. 7b), the amplitude of the seasonal variation is nearly the largest among all the regions, and this is obviously driven by snow cover. In winter, when snow cover reaches its maximum in February, the MLSE36.5 (MLSE at 36.5 GHz, hereafter) is the lowest in this region at about 0.84 (compared to 0.92 at Region A). This is because snow particles strongly scatter upwelling microwave radiation, leading to apparent decreases of MLSE. But for lower frequencies such as 10.65 GHz and 18.7 GHz, this effect is much weaker. Still the MLSE10.65 (MLSE at 10.65 GHz, hereafter) and MLSE18.7 (MLSE at 18.7 GHz, hereafter) show decreasing trends from their peak values in November to their minimums in March and April, respectively. It is possible that the melting of snow in these months greatly increases the soil moisture, resulting in minimal MLSEs. It is interesting that the frequency dependence of MLSE changes here between summer and winter (Fig. 8). In summer, MLSE increases with frequency (i.e., MLSE36.5 > MLSE18.7 > MLSE10.65); however, in winter, MLSE decreases with frequency (i.e., MLSE36.5 < MLSE18.7 < MLSE10.65) due to the previously mentioned snow scattering effect. In addition, it is found that there is a weak positive correlation between all MLSE products and NDVI in summer.
In the croplands of the northern China plain, Region C (Fig. 7c), the amplitude of the seasonal variation is also very small. However, significant signals of double-harvest per year in 2016, 2017, and 2018 can be found in NDVI and MLSE. In the months before harvest, NDVI and MLSE reach their peaks and then drop dramatically after harvest. However, the positive response of MLSE to NDVI might be partially compensated by the suppression effect from rainfall, which is heavy in summer (the crop-growing season). There is a weak increasing trend from July to December, most likely because of decreasing rainfall. As shown in Fig. 7, Region C is the only region where the MLSE is higher at a lower frequency than that at a higher frequency in all seasons (i.e., MLSE36.5 < MLSE18.7 < MLSE10.65). We speculate that the grain has a scattering effect, similar to the effect of snow particles as mentioned for Region B, and this effect decreases the upwelling microwave radiation, particularly at high frequencies.
In the boreal cropland Region D, the temporal variation is very similar to that in Region B. Also, we found large-amplitude seasonal variation due to snow. The MLSE36.5 reaches its minimal value of 0.8 or lower in February. And the MLSE10.65 and MLSE10.87 reach their minimums at the end of snow-covering season (March-April-May). The frequency dependence of MLSE reverses in winter compared to that in summer (Fig. 8). There is also a weak correlation between NDVI and MLSE.
In the grasslands of the Tibet Plateau, Region E, we found a medium amplitude of seasonal variation. MLSE is larger in fall and winter than in spring and summer. It seems this is negatively correlated to the rainfall. In summer, the increase of vegetation (NDVI) leads to a small peak of MLSE, indicating MLSE increases with vegetation here. In winter, the snow cover here is significant but is relatively lower (by 20%–40%) than in Regions B and D. Consequently, the snow does not cause remarkable suppression of MLSE36.5. In addition, differences in the snow ice microphysics also make distinct differences in the suppression effects. This is highly related to the grain size, snow water equivalent, snow depth, and frozen status. In this area, MLSE increases with increasing frequency in all seasons (Fig. 8, MLSE36.5 > MLSE18.7 > MLSE10.65).
In another grassland in Loess Plateau, Region F, the seasonal variations of NDVI, snow cover, and rainfall are all weaker than those in Region E (also grasslands). Consequently, the amplitude of MLSE seasonal variation is also weaker.
In the tropical mixed forest over Hainan Island, Region G, the seasonal variation is very weak. MLSE shows weak correlation with NDVI but increases during the dry season and decreases during the rainy season due to changes in soil moisture. Due to the impact of water in inland waterbodies and coasts, the MLSE 10.65 is very low in this region, which is also evidently demonstrated by its PDF (Fig. 8).
In the selected desert Region H, the MLSE values are the smallest among all the selected regions. The seasonal variation is medium and clearly driven by snow cover. However, the effect of snow cover on MLSE here is enhancement instead of suppression, as seen in other regions. It is possibly due to the relative difference of dielectric constants between sand grains and snow particles. There is very weak rainfall in JJA in this region, but it seems this does not cause significant impacts on the soil moisture and has no impacts on MLSE.
Interannual variations of seasonal mean MLSE from 2003 to 2019 in summer and winter are shown in Figs. 9 and 10 using both AMSR-E (e.g., Norouzi_MLSE) and FY-3B MWRI retrievals. The time series of seasonal mean MLSEs derived from the two sensors show comparable interannual variations in most regions. This again is proof that the measurements and recalibrations of MWRI are reliable and the performance of the MLSE retrieval algorithm works well.
Figure 9. Time series of seasonal mean MLSEs (at 10.65H in blue, 18.7H in orange, and 36.5H in red) in summer (June, July, and August) from 2003 to 2008 derived from AMSR-E (Norouzi_MLSE) and FY-3B MWRI from 2016 to 2019 in the eight selected regions.
Figure 10. Time series of seasonal mean MLSEs (at 10.65H as blue lines, 18.7H as orange lines, and 36.5H as red lines) in winter (December, January, and February) from 2003 to 2008 derived from AMSR-E (Norouzi_MLSE) and FY-3B MWRI from 2016 to 2019 in the eight selected regions. Overlapped pink vertical bars are associated seasonal mean snow cover (%) derived from MODIS.
It was found that in Hainan Island (Figs. 9g and 10g) there is a significant drop in MLSE, particularly at 10.65 GHz, in FY-3B retrievals in both summer and winter. It indicates that our MLSE retrieval may suffer from severe side lobe effects from open water, particularly in the coastal area. And in Region F, the grasslands in Loess Plateau, the summer MLSE shows a weak increasing trend. It is possible that such changes reflect the real land use and land changes occurring in those regions since the time series in other regions are rather stable. Of course, further in-depth investigations based on long-term data records must be done before we can draw a solid conclusion regarding the long-term trend of MLSE and its causes. In winter, the annual MLSE is heavily affected by the variations of annual snow cover. In the years with greater snow cover, MLSEs at 36.5 GHz from both sensors decrease significantly. This is evident in Region B (the boreal forest), Region D (the boreal croplands), and Region E (the grasslands over the eastern Tibet Plateau).
-
As shown in Fig. 11, the correlations between MLSEs and NDVI vary with different types of land cover and at different frequencies. In forest Region B, MLSEs tend to increase with NDVI at first when NDVI is weaker than about 0.3–0.4, and then MLSEs decrease or tend to be saturated with the further increase of NDVI. This phenomenon is also observed in Region D (the boreal croplands) and in Region E (the grasslands over the Tibet Plateau) and is particularly evident for MLSE36.5. This result is consistent with Li and Min (2013), who pointed out that the response of MLSE to vegetation water content (VWC) could be either positive or negative, depending on the partition between vegetation emission, scattering, and soil emission characteristics. In their study, MLSEs increase when VWC is small and then decrease with VWC when VWC is large. The non-monotonic dependence of MLSE on vegetation is also consistent with the theoretic calculation of Min and Lin (2006b).
Rainfall has two different impacts on MLSE. First, MLSE is expected to decrease with increasing rainfall due to its enhancement of soil moisture. Second, rainfall favors the growth of vegetation. For instance, vegetation grows with increasing rainfall in spring and generally reaches its mature stage in summer. As suggested in previous studies (Prigent et al., 2005, 2006; Shahroudi and Rossow, 2014; You et al., 2014), surface precipitation can change the soil moisture and associated emissivity within minutes and exert a long-lasting effect after the rain ends. Thus, the analysis of rainfall effects on MLSE should be carried out on the daily scale. As shown in Fig. 12, at the daily scale, there is a negative response of MLSE to increasing rainfall, especially at the lower frequency (10.65 GHz) and in non-forest areas, which indicates soil moisture changes dominate the MLSE variations. However, inherent positive correlations between rainfall and vegetation are still found, which partially compensate the response of MLSE to rainfall (soil moisture), and thus, the observed correlations between MLSE and rainfall are actually weak in most regions, as shown in Fig. 12. In Region B, the MLSE36.5 even increases with increasing rainfall. It appears that at this high frequency, MLSE reacts primarily to the vegetation increase, not to soil moisture variations, which is consistent with Prigent et al. (2006).
Figure 12. Scatter plots of monthly MLSEs against rainfall in the eight selected regions. Samples are aggregated from daily records of MLSE and GPCP rainfall around the sites in 2016.
In regions with large snow cover of up to 80% (Regions B and D, Fig. 13), the MLSE36.5 significantly decreases with increasing snow cover, and the MLSE10.65 increases with increasing snow cover. The performance of MLSE18.7 is in between. In regions with small-to-medium snow cover of less than 40% (Region C, E, and F, Fig. 13), all MLSEs show weak correlation with snow cover.
Figure 13. Scatter plots of monthly MLSEs against snow cover in six selected regions over 2015 to 2019.
However, in the desert Region H, all three MLSEs clearly increase with snow cover. This is because the electrical conductivity of dry sand is close to or even lower than that of ice particles (Ulaby et al., 1986). When the sand was covered by snow or wetted by the melting snow, the apparent MLSE increases. However, for other types of land cover, the electrical conductivity of soil under snow is relatively higher. Therefore, the snow generally leads to suppression of MLSE.
Frequencies (GHz) | ||||||
6.925 | 10.65 | 18.7 | 23.8 | 36.5 | 89.0 | |
MWRI sensitivity (K) | − | 0.5 | 0.5 | 0.5 | 0.5 | 0.8 |
AMSR-E sensitivity (K) | 0.3 | 0.6 | 0.6 | 0.6 | 0.6 | 1.1 |