Validation of downscaled 1‐km SMOS and SMAP soil moisture data in 2010–2021

Soil moisture (SM) is an important component for many applications in agriculture, hydrology, meteorology, and ecology. In past decades, passive/active microwave sensors onboard Earth observation satellites are utilized to obtain SM estimates from radiometer or radar observations. In this study, the Soil Moisture and Ocean Salinity (SMOS) Level 3 daily SM retrievals at 25‐km spatial resolution between 2010 and 2021 were downscaled through an apparent thermal inertia principle‐based algorithm. The 1‐km downscaled SMOS SM retrievals were validated by in situ measurements from 635 sites of 19 SM networks in the world, which were acquired from the International Soil Moisture Network and Texas Soil Observation Network. Additionally, the validation results of the SMOS SM products were compared with those of the Soil Moisture Active Passive (SMAP) global Level 2 enhanced SM products at 1‐km downscaled and original 9‐km resolution in 2015–2021. It shows that the downscaled SMOS SM data have an overall improved accuracy and outperform the coarse‐resolution 25‐km data, with a lower unbiased Root Mean Squared Error of 0.114 m3/m3 on average.


INTRODUCTION
In recent years, remote sensing technology has been utilized to obtain soil moisture (SM) estimates from passive microwave radiometer observations as one of important input components for many studies of land surface hydrological processes, including streamflow, surface runoff, soil infiltration, precipitation, and evapotranspiration (Dai et al., 2004;Kerr et al., 2007;Lakshmi et al., 2011Lakshmi et al., , 2018;;Wagner et al., 2007).The topsoil layer SM has an influence on the interactions between land surface processes and atmospheric boundary layer and has a direct relationship with weather and climate patterns on a global scale (Brubaker et al., 1996;Miralles et al., 2019;Pielke et al., 2001).Studies show that SM and vegetation can contribute to the development of severe storms (Al-Khaldi et al., 2019;Chang et al., 1991;Hain et al., 2011).SM observations can be assimilated into hydrological models to improve the accuracy of hydrological variables, such as surface temperature and evaporation (Beck et al., 2021;Houser et al., 1998).Finally, SM is also able to influence surface moisture gradients and sensible and latent heat flux partitioning (Engman et al., 1997;Fleckenstein et al., 2010).
In the past, the Advanced Microwave Scanning Radiometer (AMSR-E) onboard the National Aeronautics and Space Administration (NASA) Aqua satellite operated from 2002 to 2011, and it was the first satellite mission to provide standard remotely sensed SM products on a global scale (Njoku et al., 2003(Njoku et al., , 2006)).Later, the Advanced Microwave Scanning Radiometer 2 (AMSR2) instrument was designed as part of the Global Change Observation Mission-Water (GCOM-W), which was a follow-on mission of the AMSR-E satellite and started to routinely provide SM data since July 2012 (Bindlish et al., 2017;Parinussa et al., 2015;Shibata et al., 2003;Yao et al., 2021).In addition to this, the Soil Moisture and Ocean Salinity (SMOS) satellite, launched in November 2009 by the European Space Agency, is the first satellite for retrieving routine SM measurements on L-band microwave frequency (Barré et al., 2008;Kerr et al., 2012a;Kerr et al., 2012b;R. Zhang et al., 2021).The Soil Moisture Active Passive (SMAP), launched in January 2015, is NASA's first satellite to obtain global SM measurements at different resolutions including 3 km from radar (stopped operating in July 2015), 36 km from radiometer instrument, and 9 km from activepassive instruments (Chan et al., 2016(Chan et al., , 2018;;Colliander et al., 2017;Jackson et al., 2018;Kim et al., 2018;R. Zhang et al., 2019).
Remotely sensed SM products have been used in many studies.For instance, SM data were used for monitoring drought conditions from climatological and ecosystem aspects (Aghakouchak et al., 2015), evaluation of drought by calculating soil water deficit indices (Martínez-Fernández et al., 2016;Fang, Lakshmi, et al., 2021), evaluation of several applications in climate model (Loew et al., 2013), improving

Core Ideas
• Downscaled Soil Moisture and Ocean Salinity (SMOS) soil moisture (SM) retrievals have better accuracy and more spatial features than coarse SM. • Soil Moisture and Ocean Salinity (SMOS) soil moisture (SM) retrievals show a drier trend and higher spatial variability than SMAP SM. • Soil Moisture and Ocean Salinity (SMOS) soil moisture (SM) data quality is improved more from the downscaling algorithm than Soil Moisture Active Passive (SMAP).
streamflow simulation in tropical region using data assimilation techniques (Le et al., 2022;Tran et al., 2023), and evaluation of impact of SM on dust outbreaks in East Asia (Kim et al., 2015).
To retrieve SM from passive microwave brightness temperature (T B ) observations is quite challenging, since the SM retrieval algorithm needs to take account of several input parameters, such as data acquisition configurations, vegetation opacity, topography, and surface characteristics, including soil surface roughness, texture, and soil bulk density (Jackson, 1993;Jackson et al., 1991;Schmugge et al., 1994Schmugge et al., , 1998)).In response to this, the SMOS SM retrieval algorithm was developed based on an iterative approach of minimizing a cost function, which was used to determine the best set of process parameters through an iterative optimization between modeled and observed T B (Kerr et al., 2012a).Despite this, the uncertainties introduced by various input parameters and some unexpected interference during measurements, for example, anthropogenic radio frequency interference (RFI), often degrade SM retrieval accuracy.In light of these, it is necessary to perform a quality control evaluation on those radiometer-based SM retrievals before application.Previously, the SMOS SM validations have been conducted by various types of methods, including using in situ measurements, satellite SM estimates, and land surface model (LSM) outputs (Brocca et al., 2010;Jackson et al., 2011).
In the past few years, many studies focused on validating SMOS SM estimates using in situ measurements (Al Bitar et al., 2012;Delwart et al., 2008;Peischl et al., 2012;Sanchez et al., 2012).However, passive microwave SM data validation using ground-based measurements can be challenging due to the discrepancy between satellite and in situ data.First, the inconsistent spatial scale between remote sensing estimates and in situ measurements can introduce biases to the validation results.The biases can be described as overestimate/underestimate of true SM values or misrepresent SM spatial patterns on a large scale (Brocca et al., 2010).The in situ measurements usually represent a very localized region of 0.0025 m 2 , while the remote sensing estimates represent averaged values of a much larger spatial extent of around 1200 km 2 for the SMOS satellite (Jackson et al., 2011).A good solution is to use ground-based SM measurements from the core validation sites, which contain multiple well-calibrated SM stations for characterizing SM spatial variability within one satellite observation pixel (Bindlish et al., 2017).In addition, the method of analyzing the temporal stability of soil moisture is useful for identifying the soil moisture pattern and large-scale satellite SM validation (Cosh et al., 2006).Second, it is often difficult to validate remote sensing SM data by ground-based SM measurements within a long period.Furthermore, for validating SM data on a global scale, the ground-based data should represent SM conditions of different seasons, vegetation, and climate types (Jackson et al., 2012).Regarding these issues, in this study, we selected only densely distributed SM networks with long-term monitored SM measurements from different continents.
Due to the limitation of antenna diameter size of passive microwave radiometer instrument, the spatial resolution of SM retrievals is restricted to tens of kilometers which does not meet the requirement for many hydrological, meteorological, or agricultural applications on the watershed scale (Fang et al., 2014).Additionally, the mismatch issue of spatial scale between in situ SM measurements and satellite SM retrievals can be mitigated by downscaling SM data (Peng et al., 2017).This problem can be solved by downscaling the coarse-resolution satellite SM retrievals using other highresolution data sources, including radar or visible/infrared satellite observations, or LSM outputs (Bolten et al., 2003;Fang et al., 2013Fang et al., , 2018aFang et al., , 2018bFang et al., , 2019Fang et al., , 2020Fang et al., , 2022;;Fang, Lakshmi, et al., 2021;Merlin et al., 2008;Narayan et al., 2006;Peng et al., 2015;Vergopolan et al., 2020;W. Zhao et al., 2013), or applying mathematical approaches, for example, machine learning or data assimilation (Bai et al., 2019;Merlin et al., 2006;Reichle et al., 2007;Sahoo et al., 2013;W. Zhao et al., 2018).The higher spatial resolution soil moisture has been used in drought studies at both watershed (Dandridge et al., 2020) and continental scales (Fang, Kansara, et al., 2021).
The main purposes of this study are (1) to validate the global scale Level 3 multi-orbit daily SMOS SM data of the coarse (25 km) and downscaled (1 km) spatial resolution for the years 2010-2021 and (2) to compare validation results of SMOS SM products from part (1) with the Level 2 enhanced half-orbit SMAP SM products of the coarse (9 km) and downscaled (1 km) resolution from 2015 to 2021.The in situ measurements were acquired from the International Soil Moisture Network (ISMN) and Texas Soil Observation Network (TxSON).The downscaled SMOS/SMAP SM products were produced by applying an apparent thermal inertia (ATI) principle-based downscaling algorithm which was orig-inally developed by Fang et al. (2013), to the 1-km resolution gap-filled Moderate Resolution Imaging Spectroradiometer (MODIS) daily land surface temperature (LST) data set which was produced by modeling the linear relationship between MODIS LST and AMSR2 T B .The validation results will provide important knowledge for evaluating and improving the algorithms for retrieving and downscaling SMOS SM data.

DATA SETS
2.1 Remote sensing data

Soil Moisture and Ocean Salinity
The SMOS satellite was launched on November 2, 2009, and carries an L-band radiometer to observe topsoil layer SM and sea surface salinity (SSS).The Level 2 SM estimates are retrieved from Level 1 T B data, and the main objective of the SM retrieval algorithm is to provide SM estimates with an overall accuracy of 0.04 m 3 /m 3 (Kerr et al., 2012a).In this study, we used the Level 3 SMOS SM data at a 25-km standard projection grid, which were derived from the multiorbit SMOS T B observations on a satellite footprint scale of ∼40 km on average and from the algorithm which was developed by Al Bitar et al. (2017).The Level 3 SM retrieval algorithm is composed of several algorithms, including a physical model for calculating T B and a retrieval algorithm which is based on the Bayesian cost function method.The Level 3 SMOS version 300 daily SM data were acquired from the website: ftp://ext-catds-cpdc@ftp.ifremer.fr/.

Advanced Microwave Scanning Radiometer 2
The AMSR2 instrument is part of the GCOM-W, which is a follow-on mission of the AMSR-E satellite and started to routinely provide SM estimates of 1:30 AM/PM local time in July 2012 (Imaoka et al., 2010;Maeda et al., 2015;Shibata et al., 2003).The AMSR2 has a conical scan mechanism that enables to acquire day and time observations of the Earth every 2 days.The version 3 AMSR2 T B daily data at 10 km in 2012-2021 were produced by Japan Aerospace Exploration Agency's (JAXA) SM retrieval algorithm (Koike et al., 2004).They were acquired from the JAXA Global Portal System website (https://gportal.jaxa.jp/gpr/)and used for building the MODIS LST gap-filling model.

Soil Moisture Active Passive
The SMAP satellite operates on a near-polar sun synchronous orbit with a 2-to 3-day revisit frequency and overpasses of 6:00 a.m./p.m. local time, and with a native spatial resolution of ∼40 km extent covered by 3 dB beamwidth for the radiometer.The SMAP satellite is composed of a radar sensor that provided radar backscatter observations of 3 km (stopped operation in July 2015) and a radiometer that provides T B observations of 33-km nominal spatial resolution for monitoring land surface SM.Previous studies found that the RFI issue for the microwave instruments can be summarized as a substantial amount of low level (0.1-10 Kelvin) with various time-frequency characters, and it can bring biases to the passive microwave SM retrievals especially in the locations of Europe and East Asia.SMAP has an advanced microwave radiometer to provide time-frequency measurements and a complicated ground processing algorithm for detecting and mitigating the RFI issue (Piepmeier et al., 2013).The enhanced Level 2 half orbit 9-km SMAP SM version 4 product downloaded from the website (https:// nsidc.org/data/SPL2SMP_E)was used to produce the 1-km downscaled SM data for comparing with the SMOS SM data.

Moderate Resolution Imaging Spectroradiometer
The MODIS sensor is part of NASA's sun-synchronous scientific satellites Terra and Aqua, and it has 36 spectral bands covering a spectral range of 0.4-14.4μm for observing Earth surface processes.It operates with a revisit cycle of 1-2 days and various spatial resolutions from 250 m to 1 km (Tucker et al., 1979;Wan et al., 2004).The 1-km MODIS LST (MYD11A1) version 6.1 and normalized difference vegetation index (NDVI) (MYD13A2) version 6.1 data from NASA's Aqua satellite were used for implementing the downscaling model to obtain 1-km SM data.The data were downloaded from NASA's Earthdata website at (https://www.earthdata.nasa.gov/).

Advanced Very High-Resolution Radiometer
The Advanced Very High-Resolution Radiometer (AVHRR) is onboard the National Oceanic and Atmospheric Administration satellites N07-N19.The observations from AVHRR and Aqua/Terra MODIS instruments were integrated to produce the land long-term data record (LTDR) NDVI, which provides global scale climatology data since 1981 (Bédard et al., 2006;Lakshmi et al., 2001;Pedelty et al., 2007).The LTDR version 5 NDVI product in 1981-2018 at the 5-km resolution was used for building the downscaling model, and the data were downloaded from https://landweb.modaps.eosdis.nasa.gov/ltdr/.

Global Precipitation Mission
The Global Precipitation Mission (GPM) was launched in February 2014, and it was designed to provide precipitation and snow estimates on a global scale.The precipitation retrieved from the GPM satellite is at 0.1˚spatial resolution, and it has multiple temporal resolutions from 30 min to daily (Hou et al., 2014;Huffman et al., 2015).In this study, the Integrated MultisatellitE Retrievals for GPM (IMERG) Version 6 late run daily data were acquired from https://gpm.nasa.gov/data/directory for validating SMOS/SMAP SM data.

Land surface model data
The Global Land Data Assimilation System (GLDAS) integrates multiple source data including satellite and ground measurements by utilizing numerical modeling and data assimilation methods to output land surface variables for various scientific applications on, for example, meteorology, climatology, and hydrology.The latest version GLDAS V2.1 data set is available from 1948 to the present, and it has 3-h temporal resolution and 0.25˚spatial resolution (Rodell et al., 2004;J. Zhang et al., 2008).The advantages of long-term and routine availability and global spatial coverage of the GLDAS output variables could be utilized for modeling the relationship between SM/ΔLST and then for downscaling the 9-km SMAP SM.

Ground measurement data
The ISMN (https://ismn.geo.tuwien.ac.at/) hosts collections of SM ground measurements of different soil depths from 2842 stations of 71 networks by many institutes in the world, with a period of data record from 1952 to the present.The measurement units and sampling rates of different in situ measurements collected by ISMN were coordinated, and quality control was also performed (Dorigo et al., 2011;Gruber et al., 2013Gruber et al., , 2021)).In this study, the ISMN data from 18 SM networks, which include 16 dense and two sparse networks (UMBRIA and VDS), were extracted for validating SMOS and SMAP SM data sets.Detailed descriptions for all ISMN collected SM networks can be found in the following references: BIEBRZA (Musial et al., 2016), Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN) (Yang et al., 2013), FLUXNET-AMERIFLUX (Baldocchi et al., 2001), Finnish Meteorological Institute (FMI) (Ikonen et al., 2016), FR_AQUI (Al-Yaari et al., 2018), GROW (Kovács et al., 2019), Hydrological Open Air Laboratory (Blöschl et al., 2016), HOBE (Jensen et al., 2018), NAQU (Su et al., 2011), OzNet (Smith et al., 2012), Patitapu Soil Moisture Network (Hajdu et al., 2019), REMEDHUS (González-Zamora et al., 2019), Real-time In-Situ Soil Monitoring for Agriculture (RISMA) (Ojo et al., 2015), Soil Moisture Network within the Shan Dian River (T.Zhao at al., 2020), Soil moisture Sensing Controller And oPtimal Estimator (SoilSCAPE) (Moghaddam et al., 2010), UMBRIA (Brocca et al., 2011), andWegenerNet (Kirchengast et al., 2014).The extracted ground measurements for validation are from the topsoil layer at 0-to 5-cm depth and 6:00 a.m./p.m. corresponding to SMOS/SMAP overpasses from 2010 to 2021.TxSON was established in 2015 with the aim of describing the spatiotemporal variability of local-scale SM, and it has been identified as a core validation site by NASA's SMAP mission (Colliander et al., 2017).This dense monitoring network contains 40 stations that are non-randomly distributed within one 36-km Equal-Area Scalable Earth grid cell to provide hourly SM measurements at 5-, 10-, 20-, and 50-cm soil depths (Caldwell et al., 2018(Caldwell et al., , 2019)).The spatial representativeness and pronounced accuracy of SM measurements are separately guaranteed by the stratified sampling and the rigid quality flagging procedure that incorporates the coincident precipitation and SM.
Figure 1 shows the locations of SM stations from the 18 SM networks collected by the ISMN, as well as the TxSON for validating the downscaled and coarse-resolution SMOS and SMAP SM data.Four SM networks, which represent different climate types (warm temperature, semi-arid, and tundra) and land cover types (grass land, cropland, and forest), were selected for demonstrating SM spatial patterns and validation results.The SM networks are RISMA, TxSON, HOBE, and CTP-SMTMN.Table A1 shows basic information, including region, climate type, land cover, soil type, operating time, and number of stations for the 19 SM networks used for validating the SMOS and SMAP SM data sets.

Gap-filling of MODIS LST
The MODIS LST products are retrieved from visible and infrared band observations, which are usually associated with the cloud cover issue which can cause missing data and bring uncertainties to the regions around cloud and aerosol edges (Wan et al., 2008).On the other hand, the passive microwave sensors are capable of penetrating through cloud cover and could be used to provide information, while the MODIS LST data are absent.Previously, several studies focused on the relationship between passive microwave T B and LST, and the results showed a reliable range of biases in Australia (Holmes et al., 2009;Parinussa et al., 2016).Therefore, the daily MODIS LST retrievals of day/night (1:30 a.m./p.m.) can be correlated to the daily AMSR2 T B observations at .The linear regression model on the MODIS grid scale can be built as: (1) The relationship was modeled using all available daily MODIS LST and AMSR T B data between 2012 and 2021.Additionally, as Figure 2 shows, the LST-T B relationship shows significant seasonal variability especially from the slope map (Figure 2c), so each month should be modeled separately.The modeled equation was then applied to the daily AMSR T B to calculate the simulated 1-km LST G for gap-filling missing values of MODIS LST.

Downscaling algorithm
The 1-km downscaled SMOS/SMAP global SM data were produced by a downscaling algorithm developed based on the ATI principle (Fang et al., 2013).The algorithm can be described as a relationship between soil wetness, LST, and vegetation which is named the universal triangle principle (Carlson, 2007;Gillies et al., 1997.In the past, this relationship was modeled, and SM can be quantified by LST and NDVI (Mallick et al., 2009;Minacapilli et al., 2009).Therefore, the downscaling algorithm was developed on the linear regression relationship between diurnal LST change and SM corresponding to 6:00 a.m./p.m. overpasses respectively on a monthly basis by using two model output variables of the GLDAS Noah 3-hourly global data in 1981-2018: average surface skin temperature and 0-to 10-cm liquid soil moisture content.The relationship between SM  and LST difference Δ at GLDAS pixel scale can be modeled by the following equation: where  0 and  1 are the linear regression coefficients to be determined.Additionally, the  − Δ point pairs were primarily classified into 10 NDVI classes between 0 and 1 (0.1 as the interval) based on the point corresponded NDVI value, when implementing the downscaling model.This process is used accounting for the effect of vegetation modulation on the SM response to the LST change (Carlson et al., 2007;Lakshmi et al., 2013;Mallick et al., 2009).
In order to remove the discrepancy between the  − Δ model output SM  , and the SMOS/SMAP SM, a downscaling step was conducted on  , as: where  ′ ,  is the downscaled 1-km SMOS/SMAP SM pixel, modified by adding the difference between the coarseresolution SMOS/SMAP SM value Θ, and the sum of model output  , pixels within a 33-km SMOS/SMAP SM pixel domain, which is the native resolution of satellite observation.

Validation metrics
The 1-km downscaled SMOS/SMAP SM from 2010 to 2021 were validated by in situ SM measurements from 19 SM networks from the ISMN and TxSON.The following statistical metrics were calculated: coefficient of determination R 2 , root mean squared error (RMSE), unbiased RMSE (ubRMSE), bias, mean absolute error (MAE), and spatial standard deviation (SSD).The equations are as follows: where  and θ are in situ and 1-km downscaled SM data, respectively, ubRMSE is the unbiased RMSE after removing the mean bias  from RMSE, and SSD represents the spatially characterized standard deviation for all stations within one SM network.

Evaluation of the MODIS LST gap-filling model
Figure 2 shows global maps of linear regression model evaluation variables from the MODIS LST gap-filling model between AMSR2 T B and MODIS LST at the 10-km AMSR2 pixel resolution.The variables include R 2 and slope of ascending overpass (1:30 p.m.), as well as RMSE differences between ascending/descending overpasses (1:30 p.m. and 1:30 a.m.).Generally, seasonal variations for all three variables can be noted.For the R 2 maps, April and October have higher overall R 2 values worldwide.Particularly, middle-and high-latitude regions show greater variability of R 2 .The tropical regions-Amazon Rainforest, Central Africa, and Southeast Asia-consistently have relatively lower R 2 throughout the year.The main reason could be due to the data accuracy of MODIS LST in the tropical region, which is often impacted by cloud cover issues, and it may contribute to the low correlation between AMSR2 T B and MODIS LST.For the RMSE maps, higher values are found in high-latitude regions above 45˚N, as well as high-altitude regions, such as the Rocky and Andes Mountains and the Tibetan Plateau.This issue can be caused by great annual temperature variation in the abovementioned regions.From the soil hydrological point of view, these regions usually have sparse or small vegetation cover which can reduce evapotranspiration and have shallow soil with constrained water movement.Therefore, low spatial and high temporal LST variability can be found in these regions.If comparing between different months, it can be summarized that more than 50% land pixels have R 2 values >0.5 or ΔRMSE absolute values <2 K.This fact proves that the simulated LST differences from the linear regression model have reliable accuracy and can be used to gap-fill the MODIS LST product.Lastly, for the slope maps, clear seasonal variation from January to October is found especially in high-and middle-latitude regions.

4.2
The SMOS/SMAP SM maps In Figure 3, the 3-day composite 1-km downscaled SMOS/SMAP SM by the ATI downscaling algorithm are compared with the coarse-resolution 25-km SMOS and 9-km SMAP SM from August 1 to 6, 2021, on a global scale.
It can be noted that SMOS SM values are normally lower than SMAP, for example, in Central Asia and Australia.
When comparing the four SM products on a watershed scale within a 1-month period (Figure 4 and Figure A3), the abovementioned difference between SMOS/SMAP can be better observed.In addition to this, the discrepancy of spatial pattern can be noted especially in high-latitude regions, for example, Eastern Canada and Russian Far East.The spatial pattern between the downscaled and coarse-resolution SM maps for either SMOS or SMAP is consistent.Additionally, it also can be found that the two downscaled SM data show a large area of missing values, especially in middle-latitude regions.This issue is mainly due to two reasons: first, the downscaled SM values are derived from remote sensing data and LSM data sets through the LST gap-filling algorithm and the downscaling algorithm processes.In addition, the present gaps in all data sets are mainly caused by cloud cover, thick vegetation, swath gaps, or dropped low-quality retrievals.Second, the downscaled SM values may be out of the valid SM range (0-0.6 m 3 /m 3 ) especially in the regions with extremely high or low SM values and thus have to be discarded.Figure 4 shows SM maps of the downscaled and coarse-resolution SMOS and SMAP in river basins derived from a global river networking model by Lehner and Grill (2013).In Figure 4a-d, the SMOS SM maps show greater temporal variation, and they are overall drier than SMAP SM throughout the entire month.The SM spatial pattern for SMAP shows greater spatial continuity than SMOS, which may be due to better resolution for SMAP and enhanced interpolation procedure (Chan et al., 2018) on the original Level 1 SMAP T B data.Regarding the spatial feature discrepancy between the two data sets, two wetting episodes in the southeast and southwest of the watershed in Figure 4b, as well as a dry-down episode in the east of the watershed in Figure 4d can be found in the SMOS maps, which are not shown in the corresponding SMAP maps.The downscaled SMOS/SMAP maps both show greater spatial/temporal SM variability than the coarse-resolution SM maps.It is worth mentioning that the missing pixels along coastal regions are due to coarse resolution of the 25-km GLDAS data, which were used to build the ATI downscaling model.This issue is reflected in Figure 4c, which shows that the western area of the Skjern River basin has no SM estimates on the 1-km SMOS/SMAP maps, as the river basin lies on the west coast of Denmark.Additionally, the blocky effect, which means spatial discontinuity of SM between adjacent pixels, can be noticed in the downscaled maps.The downscaled SMOS maps show more blocky effects than the downscaled SMAP maps.This is mainly due to the inconsistent spatial resolution of the data sets used for building and implementing the downscaling algorithm.

Analyses of the SMOS/SMAP SM validations
Table 1 shows the averaged statistical metrics of the 19 global SM networks from validation results of the SMOS 1-and 25-km SM of ascending overpass (6:00 a.m.).It is noted that for the 1-km SMOS validations, TxSON has the highest R 2 (0.579) among all SM networks and relatively low F I G U R E 3 World maps of the three-day composite 1-km downscaled Level 3 Soil Moisture and Ocean Salinity (SMOS) and Level 2 enhanced Soil Moisture Active Passive soil moisture (SMAP SM) products by the apparent thermal inertia (ATI) downscaling algorithm, as well as the coarse resolution 25-km SMOS and 9-km SMAP SM products from August 1 to 6, 2021.ubRMSE (0.059 m 3 /m 3 ).If we compare the statistical metrics for 1 km with those for 25 km, it is found that 1-km data mostly have smaller values of both ubRMSE (ranges 0.034-0.425m 3 /m 3 ) and MAE (ranges 0.028-0.41m 3 /m 3 ) than those (ubRMSE ranges 0.037-0.46m 3 /m 3 and MAE ranges 0.034-0.443m 3 /m 3 ) for 25 km in most of the networks, except for three networks: FLUXNET-AMERIFLUX, HOBE, and UMBRIA.These results indicate an overall improvement in data accuracy of the 1-km downscaled SMOS SM over the 25-km SM.This fact can also be observed in Figure 5.In Figure 5d, the 25-km SM data points demonstrate an underestimation trend for all four stations in TxSON, while the 1-km data points are better lined up, and the corresponding linear regression best-fit lines are closer to the 1-1 line.Additionally, comparing the station-specified subfigures in Figure 5, TxSON and CTP-SMTMN show a greater improvement of the 1-km SM data accuracy than the other two networks HOBE and RISMA, as the former two networks are located in dry climate and sparse vegetation-covered regions and have less total SM variations.From the time-series analysis results of the monthly averaged SM estimates shown in Figure 6 and Figure A4 , it can be found that the SMOS 1-and 25-km SM mostly agree well with the in situ data through time, but 1-km SM have less discrepancy with in situ than 25-km SM.This fact can be better viewed from Figure 6 RISMA and TxSON stations.Furthermore, we may notice that the in situ measurements have overall less temporal variation than either the 1-or 25-km SMOS SM through all years.In addition, precipitation could be a major issue affecting the accuracy of SMOS SM data, and SM values are often biasedly estimated around rainy days.There are two reasons that precipitation can be associated with this issue: first, in situ measurements

SM network
No  and satellite observations may have different readings especially shortly after a rainfall event.In situ measurements can rapidly go down and have lower values than satellite observations.Additionally, the ground emission observed by the satellite instrument could be attenuated and scattered by the rain droplets during rain and consequently bring uncertainties to the SM retrievals (Colliander et al., 2020).
From Table 2 that shows SMOS and SMAP validations of the overlapping period 2015-2021, it can be known that the 1-km SMOS SM have smaller ubRMSE and MAE than the corresponding 25-km SMOS SM for most of SM networks, except for HOBE and OzNet.When we intercompare validation results between the SMOS and SMAP SM, there are 10 networks in total that have smaller overall ubRMSE and MAE values for the 1-km SMOS than those for the 1-km T A B L E 2 Intercomparison of averaged validation metrics: number of data points, R 2 , unbiased root mean squared error (ubRMSE), and mean absolute error (MAE) between (a) the 1-/25-km Soil Moisture and Ocean Salinity soil moisture (SMOS SM) of ascending overpass (6 a.m.),(b) the 1-/9-km Soil Moisture Active Passive (SMAP) of descending overpass (6 a.m.) from 2015 to 2021, and (c) difference of metrics between the SMOS and SMAP SM of fine/coarse resolutions, validated by in situ SM measurements from 19 SM networks in the world.SMAP.Regarding the downscaling model performance, from Table 2c, it can be summarized that there are fewer differences of ubRMSE and MAE between SMOS and SMAP for the 1-km downscaled SM than those metrics for the original SM in most of networks, except for FMI and RISMA.Furthermore, besides the five networks which have lower ubRMSE and MAE for the 25-km SMOS than those for the 9-km SMAP (FR_AQUI, HOBE, NAQU, OzNet, and VDS), there are another five networks that have lower ubRMSE and MAE for the 1-km SMOS than those metrics for the 1-km SMAP, including FLUXNET-AMERIFLUX, GROW, REMEDHUS, SoilSCAPE, and TxSON.These results convince us that the downscaling algorithm improves more on SMOS SM data quality than SMAP.In addition to this, it can also be summarized that most of the SM networks that have worse ubRMSE/MAE for the 1-km SMOS are located in Europe, such as WegenerNet, UMBRIA, and BIEBRZA, where there is greater vegetation cover.Therefore, it can be concluded that the 1-km downscaled SMOS SM have similar data accuracy as the 1-km SMAP SM, and the downscaling algorithm performs better in the regions with uniform land cover types and low vegetation cover.

SM
From Table 3, by comparing the averaged SSD for in situ, 1-and 25-km SMOS SM data sets through each network, it can be summarized that the SSD of in situ SM, which ranges from 0.038 to 0.266 m 3 /m 3 , is generally greater than either 1 or 25 km for most of the networks, except for UMBRIA and WegenerNet.This is mainly due to the different scales of which in situ measurements and SMOS SM represent.In addition to this, the SSD of 1 km, which ranges from 0.002 to 0.106 m 3 /m 3 , is greater than that of 25 km, which ranges from 0.001 to 0.066 m 3 /m 3 .In addition, the SSD of 1 km is closer to the corresponding SSD of in situ data for most of the stations.The only exceptions we can find are OzNet and SoilSCAPE.This proves that 1-km downscaled SM data can provide more SM spatial information in detail than 25-km SM on a fine scale.However, the exceptional cases may be caused by an insufficient number of in situ stations within one SMOS grid for representing SM spatial variability on the SMOS scale.On the other hand, the spatial validation was also conducted on the densely distributed in situ SM networks (as Figure 1 shows), and the in situ measurement points within one 25-km SMOS pixel boundaries were compared with the corresponding 1-or 25-km SMOS SM values.Table 4 and Appendix Table A2 show the averaged and daily validation metrics of the 1-/25-km SMOS SM from four SMOS pixels of three SM networks, respectively.It can be summarized that the ubRMSE of the 1-km SMOS (ranges 0.055-0.088m 3 /m 3 ) is smaller than that of the 25km SMOS (ranges 0.062-0.116m 3 /m 3 ) for all four SMOS pixels.In addition, MAE and bias are also improved for the 1-km SMOS in three SMOS pixels, except for TxSON #2.These results agree with the abovementioned temporal val-idation results well and also ascertain that the downscaled 1-km SMOS SM have a better overall accuracy than the 25-km SMOS SM.

CONCLUSIONS
In this study, an ATI principle-based downscaling algorithm was applied to generate the global 1-km daily SMOS SM product in 2010-2021, by using the 1-km Aqua MODIS LST data which were gap-filled through the linear regression modeling between MODIS LST and AMSR2 T B data.The 1-km SMOS SM data of ascending overpass (6:00 a.m.) were validated by ISMN and TxSON ground SM measurements from 19 SM networks in the world, as well as compared with validation results from the 1-km downscaled and 9-km SMAP

SM network
No SM from the downscaling algorithm.The results show that for the temporal validations, the overall ubRMSE is 0.104 m 3 /m 3 , and MAE is 0.097 m 3 /m 3 .On the other hand, for the spatial validation results within SMOS pixel from selected SM networks, ubRMSE ranges from 0.055 to 0.088 m 3 /m 3 , and MAE ranges from 0.045 to 0.077 m 3 /m 3 for the 1-km SMOS SM.Both temporal/spatial validations for the 1-km SMOS SM generally outperform those for the 25-km SM.
Regarding the intercomparison of validation results between SMOS and SMAP SM in 2015-2021, the data accuracy for the SMOS SM is improved more than SMAP from coarse to downscaled resolution, especially in the regions with less vegetation, which evidences that the downscaling algorithm can be recommended for fine-scale studies using SMOS SM data.Additionally, it is found that precipitation may contribute to the biased estimation of SMOS SM, and this issue is reduced on the 1-km downscaled SMOS SM to a certain extent.Finally, the 1-km SMOS SM data show overall greater SSD values than the 25-km SMOS SM.The SSD values for 1-km SM are closer to those for in situ measurements in most of SM networks than those for the 25-km SM, which suggests that the 1-km SM can provide more detailed SM spatial information.Several limitations regarding the downscaled SMOS SM data should be considered: first, the AMSR2 (10 km) T B data were used for building the model to fill data gaps of the MODIS LST at 1 km, while the downscaling model was built at 25-km resolution by the GLDAS output variables.The discrepancy in spatial resolution, sensing depth, and data measuring instrument may bring uncertainties to the downscaling algorithm building and implementation.Second, the data accuracy of SM retrievals could be affected by extreme hydrologic conditions and vegetation cover.The vegetation effect on SM retrievals and the downscaling process should be considered and calibrated beforehand.Furthermore, as the downscaled SM data are derived from multiple data sources including the original SMOS SM data and remote sensing/LSM data which have different sources of data unavailability, the downscaled SM data have less spatial coverage than the original data.Lastly, the number and representativeness of in situ stations within the SMOS grid extent should be taken into account for validation.

F
I G U R E 1 World map shows the locations of soil moisture (SM) ground stations from 19 networks which are used for validating the Soil Moisture and Ocean Salinity/Soil Moisture Active Passive (SMOS/SMAP) SM estimates.The four satellite maps show land covers of the selected four SM networks used for demonstrating correlation scatter plots and time-series plots.The networks are Real-time In-Situ Soil Monitoring for Agriculture (RISMA), Texas Soil Observation Network (TxSON), HOBE, and Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN).The 25-km SMOS grids on satellite maps are outlined in yellow color.FMI, Finnish Meteorological Institute; HOAL, Hydrological Open Air Laboratory; PTSMN, Patitapu Soil Moisture Network; SMN-SDR, Soil Moisture Network within the Shan Dian River; SoilSCAPE, Soil moisture Sensing Controller And oPtimal Estimator.

F
Three-day composite Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive soil moisture (SMAP SM) maps of coarse and downscaled resolution from August 2021 in four river basins: (a) eastern part of Assiniboine-Red River basin, corresponding to Real-time In-Situ Soil Monitoring for Agriculture (RISMA), (b) central part of Middle Colorado River basin, corresponding to Texas Soil Observation Network (TxSON), (c) Skjern River basin, corresponding to HOBE, and (d) upper part of Nu-Salween River basin, corresponding to Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN).F I G U R E 4 Continued T A B L E 1 Averaged validation metrics: number of stations, R 2 , unbiased root mean squared error (ubRMSE), and mean absolute error (MAE) of the 1-and 25-km Soil Moisture and Ocean Salinity soil moisture (SMOS SM) of ascending overpass (6 a.m.) from 2010 to 2021 by in situ SM measurements from 19 networks in the world.

F I G U R E 5
Validation scatter plots of the 1-and 25-km Soil Moisture and Ocean Salinity (SMOS) comparing with in situ soil moisture (SM) measurements of ascending overpass (6:00 a.m.) in 2010-2021 from four networks: (a) Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN), (b) HOBE, (c) Real-time In-Situ Soil Monitoring for Agriculture (RISMA), and (d) Texas Soil Observation Network (TxSON).

F
Time-series graphs of the monthly averaged 1-and 25-km Soil Moisture and Ocean Salinity (SMOS) and in situ soil moisture (SM) measurements of ascending overpass (6:00 a.m.) in 2010-2021 from four networks: Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN), HOBE, Real-time In-Situ Soil Monitoring for Agriculture (RISMA), and Texas Soil Observation Network (TxSON).The blue bars on top of graphs are the monthly accumulated Global Precipitation Mission (GPM) precipitation.

R
Bin Fang: Conceptualization; data curation; methodology; writing-original draft.Venkataraman Lakshmi: Conceptualization; formal analysis; funding acquisition; investigation; methodology; project administration; supervision; writing-review and editing.Runze Zhang: Validation; writing-review and editing.A C K N O W L E D G M E N T S The authors sincerely acknowledge the funding from the NASA Terrestrial Hydrology-NASA Award Number 80NSSC19K0993, Program Manager Dr. Jared Entin.O R C I D Bin Fang https://orcid.org/0000-0002-0448-7616Descriptions of the 19 soil moisture (SM) networks used for validating the Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive (SMAP) SM data.

T a b l e A 2
Spatial validation metrics: unbiased root mean squared error (ubRMSE), mean absolute error (MAE) and bias of the 1-and 25-km Soil Moisture and Ocean Salinity soil moisture (SMOS SM) of ascending overpass (6 a.m.) validated by in situ SM measurements from four SMOS pixels of three densely distributed networks: (a) Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN), (b) HOBE, (c) Texas Soil Observation Network-1 (TxSON-1), and (d) TxSON-2, from June to September 2016-2019.

F i g u r e A 3
Three-day composite Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive soil moisture (SMAP SM) maps of coarse and downscaled resolution from August 2021 in four river basins: (a) eastern part of Assiniboine-Red River basin, corresponding to Real-time In-Situ Soil Monitoring for Agriculture (RISMA); (b) central part of Middle Colorado River basin, corresponding to Texas Soil Observation Network (TxSON); (c) Skjern River basin, corresponding to HOBE; and (d) upper part of Nu-Salween River basin, corresponding to Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN).

F i g u r e A 4
Time-series graphs of the monthly averaged 1-and 25-km Soil Moisture and Ocean Salinity (SMOS) and in situ soil moisture (SM) measurements of ascending overpass (6:00 a.m.) in 2010-2021 from four networks: (a) Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network (CTP-SMTMN), (b) HOBE, (c) Real-time In-Situ Soil Monitoring for Agriculture (RISMA), and (d) Texas Soil Observation Network (TxSON).The blue bars on top of graphs are the monthly accumulated GPM precipitation.
Finnish Meteorological Institute; HOAL, Hydrological Open Air Laboratory; RISMA, Real-time In-Situ Soil Monitoring for Agriculture; PTSMN, Patitapu Soil Moisture Network; SMN-SDR, Soil Moisture Network within the Shan Dian River; SoilSCAPE, Soil moisture Sensing Controller And oPtimal Estimator.

. of points in pixel SMOS 1 km SMOS 25 km ubRMSE MAE Bias ubRMSE MAE Bias
: CTP-SMTMN, Central Tibetan Plateau Soil Moisture and Temperature Monitoring Network; FMI, Finnish Meteorological Institute; HOAL, Hydrological Open Air Laboratory; PTSMN, Patitapu Soil Moisture Network; RISMA, Real-time In-Situ Soil Monitoring for Agriculture; SMN-SDR, Soil Moisture Network within the Shan Dian River; SoilSCAPE, Soil moisture Sensing Controller And oPtimal Estimator; TxSON, Texas Soil Observation Network. Abbreviations