A robust regional soil moisture estimation method based on spatiotemporal two-stage data-driven models

Root-zone soil moisture (RZSM) plays a key role in the hydrologic cycle and regulates water–heat exchange. Although site observations can provide soil profile moisture measurements, they have a restricted representation. Satellites can determine soil moisture on a large scale, yet the depth of detection is limited. RZSM can be estimated on a large scale using the soil moisture analytical relationship (SMAR) and surface soil moisture (SSM). However, the applicability of the SMAR to different deep-root zones and covariate sources is unclear. This paper investigates the applicability of the SMAR in the Shandian River Basin, upstream of the Luan River in China, by combining site and regional soil moisture, soil properties, and meteorological data. In particular, we first compared the estimation results of the SMAR at different depths (10–20 cm; 10–50 cm) and using covariates from different sources (dataset, SMAR-P1; literature, SMAR-P2) at the site in order to generate SMAR calibration parameters. The parameters were then regionalized based on multiple linear regression by combining the SMAR-P1, SMAR-P2, and SMAR calibration parameters in the 10–50 cm root zone. Finally, the Shandian River RZSM was estimated using regional surface soil moisture and the aforementioned regionalized parameters. At the site scale, diffusion coefficient b obtained in the 10–20 cm root zone at the same depth as the surface layer exceeded the upper limit of the SMAR by one. This is not fit an environment within the site context, and thus the SMAR is not applicable at this particular depth. The opposite is observed for the 10–50 cm root zone. In addition, SMAR-P1 (RMSE = 0.02) outperformed SMAR-P2 (RMSE = 0.04) in the estimation of the RZSM at 10–50 cm. Parameter regionalization analysis revealed the failure of SMAR-P2 to pass the significance test (p > 0.05) for building a multivariate linear model, while SMAR-P1 successfully passed the significance test (p < 0.05) and finished the parameter regionalization process. The median RMSE and median R2adj of the regional RZSM results were determined as 0.12 and 0.3, respectively. The regional RZSM agrees with the spatial trend of the Shandian River. This study examines the suitability of the SMAR model in varying deep-root zones and with diverse covariate sources. The results provide a crucial basis for future utilization of the SMAR.

Abstract. Soil moisture influences numerous environmental processes occurring over large spatial and temporal scales. It profoundly influences the hydrological and meteorological activity together with climate predictions and hazard analysis. Space-borne sensors are capable of retrieving the surface soil moisture over a region on a regular basis. Latent heat measurements of soil, reflectance based methods, microwave measurements and synergistic approaches are some of the techniques used since long for providing soil moisture estimates over regional and global scales. Due to the dynamic interaction of soil with crops, retrieval of surface soil moisture is always challenging. This paper gives a brief overview of advance in soil moisture retrieval techniques, and an attempt to generate surface soil moisture from fine-resolution satellite remote sensing data. The optical remote sensing explores the linear relationship between land surface reflectance and soil moisture content, and through development of empirical spectral vegetation indices. Another way to estimate soil moisture emerged by measuring amplitude of diurnal temperature, which is closely related to thermal conductivity and heat capacity of soil. Emergence of radiometric satellite measurements at fine resolution has reached at a higher level of technology these days. Microwave remote sensing techniques have a long legacy of providing surface soil moisture estimates with reasonable accuracy. The SMOS (Soil Moisture and Ocean Salinity) and SMAP (Soil Moisture Passive and Active) missions launched in 2009 and 2015 respectively, are completely dedicated for providing soil moisture at global scale with a spatial resolution of 35 km & 3–40 km. These soil moisture products, however, provides data at highly coarser spatial resolution. The launch of Sentinels gave insight by providing active radar and optical data at higher resolution (∼10 m). Sentinel-1 is the first SAR (Synthetic Aperture Radar) constellation having 6-day revisit time providing data in C-band with dual polarisations. However, no algorithm or methodology is available to generate surface soil moisture product at a finer resolution from dual polarisations. Sentinel-1 data has been used to generate regional surface soil moisture image through modelling. The same has been also used for generating surface soil moisture map of IARI farm at New Delhi. Dubois, a bare surface model, was tested for its suitability for surface soil moisture retrieval of the farm. In addition, radar- based Soil moisture (SM) proxy method was used over Sentinel-1 data for the month of July 2018, and validated through actual surface soil moisture (gravimetric) measurements. Results were satisfactory for a range of 4–16 m3 m−3 of soil moisture, with coefficient of determination (R2) as 0.45, RMSE of 2.35 and a p-value of 0.005. However, over a higher range of soil moisture (21–33 m3 m−3), which occurred after the rainfall, the R2 value reduced to 0.22 with larger RMSE. Results suggested that SM-proxy approach might work well for a limited range (drier part) of soil moisture content, and not for the wet soil.

Accurate measurement and monitoring of surface and subsurface soil moisture is essential for understanding hydrological processes, crop growth modeling, crop water requirement, and climate studies. Accurate measurement of the soil moisture content (SMC) in the root zone is essential for precise irrigation authority and plant water stress evaluation. However, the existing passive microwave satellite missions, Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive (SMAP), that operate at L-band, can only estimate the top 5 cm of soil moisture. Microwave remote sensing has proven to be a valuable tool for non-invasive soil moisture estimation. This research aims to investigate and develop a methodology for estimating surface and subsurface soil moisture using microwave data from Sentinel-1. The study was conducted to establish the relationship between surface & the backscatter coefficient derived using the Sentinel-1 SAR microwave remote sensing satellite imagery, and relationship between surface and subsurface soil moisture at different depths, in the Godhra region. Two seasons namely summer (Zaid) and monsoon (Kharif) were taken into consideration to build up the relationship between surface soil moisture and co-polarization backscatter coefficient ( For the summer (Zaid) and monsoon (Kharif) seasons, the co-polarization backscatter coefficient ( and surface soil moisture (0-5, cm) were found to have a correlation in terms of R2 as 0.91 and 0.90, respectively. The study explores the relationship between microwave signals and surface soil moisture content (0-5, cm) and then the relationship between surface soil moisture and soil moisture at various depths were also modeled thereby contributing to improved soil moisture estimation techniques and applications. The value of the coefficient of determination (R2) of surface soil moisture (0-5, cm) to subsurface soil moisture at 6-20 cm, 21-40 cm, and 41-60 cm depths were found to be 0.60, 0.51, and 0.46, respectively, in the summer (Zaid) season. The value of the coefficient of determination (R2) of surface soil moisture (0-5, cm) to subsurface soil moisture at 6-20 cm, 21-40 cm, 41-60 cm, 61-80 cm, and 81-100 cm depths were found to be 0.83, 0.61, 0.51, 0.26, and 0.13, respectively. According to the study, it is observed that the relationship between co-polarization backscatter coefficient ( and soil moisture weakens as the depth of soil moisture increases. Overall, the regression models developed between the co-polarization backscatter coefficient ( and surface soil moisture showed very good results, whereas the regression models developed between the surface soil moisture and soil moisture at various depths showed reasonably acceptable results up to the depth of 60 cm. The findings in the present study suggest that Sentinel-1A C-band SAR data can be used to estimate surface soil moisture. It is also shown in this study that the surface soil moisture can be correlated with the subsurface soil moisture up to the depth of 60 cm, satisfactorily using regression equations.

The difficulty of accurate and large-scale measurement for surface parameters limits the regional surface soil moisture (SSM) estimation using synthetic aperture radar (SAR). Moreover, the coarse resolution of soil moisture products generated by existing methods, which fuse SAR and passive microwave products, cannot fully satisfy the requirement of specific regional applications. To solve this problem, an SAR-optical data fusion method for soil moisture estimation (SOFSME) based on a cascade neural network is proposed in this study. SOFSME obtains surface parameters from historical soil moisture images and related environmental images to estimate a SSM image with high resolution at large scale from Sentinel-1A C-band SAR data. Validation experiments in single and multiple land-use type areas showed that the SOFSME performed best on bare soil areas with a median root mean square error of 0.0203. The median universal image quality index of estimated soil moisture image was 0.1454, which was better for single cropland areas than multi-land-use type areas. The Pearson correlation coefficient showed a median value of 0.7645 in both experiments. These results showed that the SOFSME had high accuracy, availability, and stability in regional soil moisture estimation. Compared with existing methods, the SOFSME can provide high-quality soil moisture images and does not directly depend on field measurement data. Thus, the proposed SOFSME method is of great value for high-resolution soil moisture estimation in more regional applications.

Improving remotely sensed soil moisture estimates requires calibration and validation from ground-based observations obtained from established monitoring networks. Network sites are often installed at the edges of fields (in grass strips), and it is unknown if the soil moisture conditions at the network sites are similar to those observed within the fields. Intensive field campaigns, that include extensive spatial sampling of soil moisture, can be used as a basis for comparison for network sites. This study utilized data from the Canadian Experiment for Soil Moisture in 2010 (CanEx-SM10). Regional mean soil moisture (at the scale required for passive microwave remote sensing) obtained from the network sites (32 in total) was compared to the mean soil moisture obtained from field locations (55–60 fields) within the same region for the 6 days of the field campaign. The mean difference between the regional mean network soil moisture and the regional mean field soil moisture was < 0.04 m3 m−3 for each day of the campaign. A bootstrapping technique, which randomly sampled the network data, determined that the regional field mean soil moisture fell within the 95% confidence interval for the network data for all days and resulted in a root mean square error (RMSE) between the network and the regional field soil moisture of < 0.03 m3 m−3. Thiessen polygons were used as an upscaling technique to determine the regional-scale soil moisture resulting from network and manual field measurements. The results indicated that the difference between the regional-scale soil moisture from the network versus the field measurements was < 0.041 m3 m−3 for all sampling days. A Monte Carlo analysis indicated that 25 of the network stations (within a region of approximately 1600 km2) would be required in order for the network mean to be within 0.04 m3 m−3 of the field mean soil moisture with 95% confidence.

Investigating the spatiotemporal changes of regional surface soil moisture in responses to climate and land cover changes is vital for understanding the underlying mechanisms of hydrological processes. While previous studies mainly attributed the causes of soil moisture changes to climatic factors, few took land cover into consideration. We analyzed the seasonal‐differentiated effects of climate and land cover changes on surface soil moisture of China's 80 drainage basins using the Essential Climate Variable Soil Moisture product spanning the 1979–2010 period. The low–low spatial clusters of annual (Jan–Dec), warm‐season (Apr–Sept) and cool‐season (Oct–Dec and Jan–Mar) surface soil moisture have spread from Northwest China to northeast and even Central China during the past three decades. In cool seasons, significant decreasing trends of surface soil moisture in most drainage basins of Northwest, Northeast, Central, and South China were detected. But in warm seasons, the surface soil moisture in Central and South China showed significant increasing trends. Precipitation, potential evapotranspiration and the ratio of the two are three driving forces for warm‐season surface soil moisture changes in Northeast and Northwest China. These climatic factors are main contributors to the declining trends of cool‐season surface soil moisture in Central and South China. Land cover changes showed to be the major factor driving the significant decreasing trends of cool‐season surface soil moisture in Northwest, Central, and South China. Due to vegetation's self‐shading effect, the large‐scale reforestation and vegetation growth are believed as main causes of the increasing trends of warm‐season surface soil moisture in Central and South China.

Soil moisture (SM) is an important driver for forest ecosystems, creating a need for globally extensive SM information that can only be achieved with satellite-based sensors and/or process-based model. However, the reliability of remotely sensed or modeled SM data in forests is poorly understood due to a lack of suitable validation sites and interference with remote sensing caused by vegetation water content. Here we examine three multi-year SM products: (i) remotely sensed surface (0-5 cm) SM from combined Soil Moisture Active Passive (SMAP) and Sentinel-1 observations (SMAP/Sentinel), (ii) the SMAP Level-4 surface (0-5 cm) and root-zone (0-1 m) SM data assimilation product (SMAP-L4), and (iii) simulated surface (0-10 cm) and root-zone (0-1 m) SM from the North American Land Data Assimilation System (NLDAS). These estimates were compared with in situ measurements from 39 National Ecological Observatory Network sites throughout the US. At 21 unforested sites, the performance of the three products was similar for surface SM, and all three were able to track temporal changes in surface SM. The performance of the three products declined at 18 forested sites; however, while the performance declined modestly for SMAP-L4 and NLDAS, SMAP/Sentinel performance declined so much that it was largely unable to track changes in surface SM. The SMAP-L4 and NLDAS products also reliably captured temporal changes in root-zone SM at both forested and unforested sites. Our findings indicate that both SMAP-L4 and NLDAS can be used to track surface and root-zone SM changes in forests (unbiased RMSD: 0.03-0.06 m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> ).

In current upscaling of in situ surface soil moisture practices, commonly used novel statistical or machine learning-based regression models combined with remote sensing data show some advantages in accurately capturing the satellite footprint scale of specific local or regional surface soil moisture. However, the performance of most models is largely determined by the size of the training data and the limited generalization ability to accomplish correlation extraction in regression models, which are unsuitable for larger scale practices. In this paper, a deep learning model was proposed to estimate soil moisture on a national scale. The deep learning model has the advantage of representing nonlinearities and modeling complex relationships from large-scale data. To illustrate the deep learning model for soil moisture estimation, the croplands of China were selected as the study area, and four years of Visible Infrared Imaging Radiometer Suite (VIIRS) raw data records (RDR) were used as input parameters, then the models were trained and soil moisture estimates were obtained. Results demonstrate that the estimated models captured the complex relationship between the remote sensing variables and in situ surface soil moisture with an adjusted coefficient of determination of R ¯ 2 = 0.9875 and a root mean square error (RMSE) of 0.0084 in China. These results were more accurate than the Soil Moisture Active Passive (SMAP) active radar soil moisture products and the Global Land data assimilation system (GLDAS) 0–10 cm depth soil moisture data. Our study suggests that deep learning model have potential for operational applications of upscaling in situ surface soil moisture data at the national scale.

Successful monitoring of soil moisture dynamics at high spatio-temporal resolutions globally is hampered by the heterogeneity of soil hydraulic properties in space and complex interactions between water and the environmental variables that control it. Current soil moisture monitoring schemes via in situ station networks are sparsely distributed while remote sensing satellite soil moisture maps have a very coarse spatial resolution. In this study, an empirical surface soil moisture (SSM) model was established via fusion of in situ continental and regional scale soil moisture networks, remote sensing data (SMAP and Sentinel-1) and high-resolution land surface parameters (e.g. soil texture, terrain) using a quantile random forest (QRF) algorithm. The model had a spatial resolution of 100 m and performed well under cultivated, herbaceous, forest, and shrub soils (overall R2 = 0.524, RMSE = 0.07 m3 m-3). It has a relatively good transferability at the regional scale among different soil moisture networks (mean RMSE = 0.08–0.10 m3 m-3). The global model was applied to map SSM dynamics at 30–100 m across a field-scale soil moisture network (TERENO-Wüstebach) and an 80-ha cultivated cropland in Wisconsin, USA. Without the use of local training data, the model was able to delineate the variations in SSM at the field scale but contained large bias. With the addition of 10% local training datasets (“spiking”), the bias of the model was significantly reduced. The QRF model was relatively insensitive to the resolution of Sentinel-1 data but was affected by the resolution and accuracy of soil maps. It was concluded that the empirical model has the potential to be applied elsewhere across the globe to map SSM at the regional to field scales for research and applications. Future research is required to improve the performance of the model by incorporating more field-scale soil moisture sensor networks and assimilation with process-based models.

Generating surface soil moisture at 30 m spatial resolution using both data fusion and machine learning toward better water resources management at the field scale

Abstract. The current intensive use of agricultural land is affecting the land quality and contributes to climate change. Feeding the world's growing population under changing climatic conditions demands a global transition to more sustainable agricultural systems. This requires efficient models and data to monitor land cultivation practices at the field to global scale. This study outlines a spatially distributed version of the field-scale crop model AquaCrop version 6.1 to simulate agricultural biomass production and soil moisture variability over Europe at a relatively fine resolution of 30 arcsec (∼1 km). A highly efficient parallel processing system is implemented to run the model regionally with global meteorological input data from the Modern-Era Retrospective analysis for Research and Applications version 2 (MERRA-2), soil textural information from the Harmonized World Soil Database version 1.2 (HWSDv1.2), and generic crop information. The setup with a generic crop is chosen as a baseline for a future satellite-based data assimilation system. The relative temporal variability in daily crop biomass production is evaluated with the Copernicus Global Land Service dry matter productivity (CGLS-DMP) data. Surface soil moisture is compared against NASA Soil Moisture Active–Passive surface soil moisture (SMAP-SSM) retrievals, the Copernicus Global Land Service surface soil moisture (CGLS-SSM) product derived from Sentinel-1, and in situ data from the International Soil Moisture Network (ISMN). Over central Europe, the regional AquaCrop model is able to capture the temporal variability in both biomass production and soil moisture, with a spatial mean temporal correlation of 0.8 (CGLS-DMP), 0.74 (SMAP-SSM), and 0.52 (CGLS-SSM). The higher performance when evaluating with SMAP-SSM compared to Sentinel-1 CGLS-SSM is largely due to the lower quality of CGLS-SSM satellite retrievals under growing vegetation. The regional model further captures the short-term and inter-annual variability, with a mean anomaly correlation of 0.46 for daily biomass and mean anomaly correlations of 0.65 (SMAP-SSM) and 0.50 (CGLS-SSM) for soil moisture. It is shown that soil textural characteristics and irrigated areas influence the model performance. Overall, the regional AquaCrop model adequately simulates crop production and soil moisture and provides a suitable setup for subsequent satellite-based data assimilation.

Regional quantification of energy and water balance fluxes depends inevitably on the estimation of surface and rootzone soil moisture. The simulation of soil moisture depends on the soil retention characteristics, which are difficult to estimate at a regional scale. Thus, the present study proposes a new method to estimate high-resolution Soil Hydraulic Parameters (SHPs) which in turn help to provide high-resolution (spatial and temporal) rootzone soil moisture (RZSM) products. The study is divided into three phases—(I) involves the estimation of finer surface soil moisture (1 km) from the coarse resolution satellite soil moisture. The algorithm utilizes MODIS 1 km Land Surface Temperature (LST) and 1 km Normalized difference vegetation Index (NDVI) for downscaling 25 km C-band derived soil moisture from AMSR-2 to 1 km surface soil moisture product. At one of the test sites, soil moisture is continuously monitored at 5, 20, and 50 cm depth, while at 44 test sites data were collected randomly for validation. The temporal and spatial correlation for the downscaled product was 70% and 83%, respectively. (II) In the second phase, downscaled soil moisture product is utilized to inversely estimate the SHPs for the van Genuchten model (1980) at 1 km resolution. The numerical experiments were conducted to understand the impact of homogeneous SHPs as compared to the three-layered parameterization of the soil profile. It was seen that the SHPs estimated using the downscaled soil moisture (I-d experiment) performed with similar efficiency as compared to SHPs estimated from the in-situ soil moisture data (I-b experiment) in simulating the soil moisture. The normalized root mean square error (nRMSE) for the two treatments was 0.37 and 0.34, respectively. It was also noted that nRMSE for the treatment with the utilization of default SHPs (I-a) and AMSR-2 soil moisture (I-c) were found to be 0.50 and 0.43, respectively. (III) Finally, the derived SHPs were used to simulate both surface soil moisture and RZSM. The final product, RZSM which is the daily 1 km product also showed a nearly 80% correlation at the test site. The estimated SHPs are seen to improve the mean NSE from 0.10 (I-a experiment) to 0.50 (I-d experiment) for the surface soil moisture simulation. The mean nRMSE for the same was found to improve from 0.50 to 0.31.

Surface soil moisture is a key variable to describe water and energy exchanges at the surface/atm interface and measure drought and aridification. The Ts-NDVI space is an effective method to monitor regional surface soil moisture status. Due to the disturbance of multiple factors, the established dry or wet boundary with monotemporal remote sensing data is unstable. This paper developed a Ts-NDVI triangle space with MODIS NDVI dataset to monitor soil moisture in the Mongolian Plateau in 2000–2012. Based on the temperature vegetation dryness index (TVDI), the spatiotemporal variations of drought were studied. The results indicated that (1) the general Ts-NDVI space method is an effective way to monitor regional soil moisture. However, if the single time space shows perfect structure, there would be no differences between the inverted results of the single time space and the general space. (2) The TVDI calculated in the paper is expected to show the water deficit for the region from low (bare soil) to high (full vegetation cover) NDVI values, and it is found to be in close negative agreement with precipitation and soil moisture; changes in the TVDI are dependent on the water status in the study area. (3) In the Mongolian Plateau, TVDI presented a zonal distribution with changes in Land Use/Land Cover types, vegetation cover, and latitude. Drought was serious in bare land, construction land, and grassland. Drought was widely spread throughout the Mongolian Plateau, and there was aridification in the study period. Vegetation degradation, overgrazing, and climate warming could be considered as the main reasons.

Estimation of root zone soil moisture and surface fluxes partitioning using near surface soil moisture measurements

&amp;lt;p&amp;gt;Current in situ soil moisture monitoring networks are sparsely distributed while remote sensing satellite soil moisture maps have a very coarse spatial resolution. In this study, an empirical global surface soil moisture (SSM) model was established via fusion of in situ continental and regional scale soil moisture networks, remote sensing data (SMAP and Sentinel-1) and high-resolution land surface parameters (e.g., soil texture, terrain) using a quantile random forest (QRF) algorithm. The model had a spatial resolution of 100m and performed moderately well under cultivated, herbaceous, forest, and shrub soils (R&amp;lt;sup&amp;gt;2&amp;lt;/sup&amp;gt; = 0.524, RMSE = 0.07 m&amp;lt;sup&amp;gt;3&amp;lt;/sup&amp;gt; m&amp;lt;sup&amp;gt;&amp;amp;#8722;3&amp;lt;/sup&amp;gt;). It has a relatively good transferability at the regional scale among different continental and regional networks (mean RMSE = 0.08&amp;amp;#8211;0.10 m&amp;lt;sup&amp;gt;3&amp;lt;/sup&amp;gt; m&amp;lt;sup&amp;gt;&amp;amp;#8722;3&amp;lt;/sup&amp;gt;). The global model was then applied to map SSM dynamics at 30&amp;amp;#8211;100m across a field-scale network (TERENO-W&amp;amp;#252;stebach) in Germany and an 80-ha irrigated cropland in Wisconsin, USA. Without local training data, the model was able to delineate the variations in SSM at the field scale but contained large bias. With the addition of 10% local training datasets (&amp;amp;#8220;spiking&amp;amp;#8221;), the bias of the model was significantly reduced. The QRF model was also affected by the resolution and accuracy of soil maps. It was concluded that the empirical model has the potential to be applied elsewhere across the globe to map SSM at the regional to field scales for research and applications. Future research is required to improve the performance of the model by incorporating more field-scale soil moisture sensor networks and high-resolution soil maps as well as assimilation with process-based water flow models.&amp;lt;/p&amp;gt;