Soil moisture monitoring by downscaling of remote sensing products using LST/VI space derived from MODIS products

Soil moisture (SM) is a major variable in the climate system, which controls the exchange of water, energy, and carbon flux between the earth's surface and the atmosphere (Ochsner et al. 2013). In global and regional water cycles, soil moisture is also one of the key parameters related to evapotranspiration, runoff occurrence, soil permeability, and groundwater recharge (Qiu et al. 2016). This ‘soil moisture’ is the water on the upper level of the soil profile and in the root zone or vadose area. Surface SM has received considerable attention for improving the predictive skills of runoff models that target flood risk prediction and/or water resource management (Entekhabi et al. 1999). For efficient irrigation management and planning, a consistent estimate of SM from agricultural fields is required (Glenn et al. 2011; Lorite et al. 2012; Srivastava et al. 2013b). Despite the importance of surface SM information, measured SM is present in only a limited number of field sites worldwide. Many factors affect the spatial distribution of SM, such as topographic changes, soil types, vegetation, climate, and aquifer water depth (Fernández-Prieto et al. 2012). On the other hand, because SM is highly variable in terms of space and time, in situ observations of SM, such as probe observations or gravimetric measurements, do not provide accurate and comprehensive information on the temporal and spatial distribution of SM and therefore are inappropriate for regional or global applications. Therefore, large-scale monitoring of SM can only rely on space-based remote sensing (Srivastava et al. 2013a, 2014). Although there are different types of remote sensing techniques for SM retrieval depending on the different electromagnetic spectra used, such as thermal, microwave, and visible infrared spectra, the accuracy of the prediction is ultimately determined by the soil moisture model.

The four dominant SM remote sensing methods based on spectrum characteristics are described below (Walker 1999; Vicente et al. 2004)

• Passive microwave (e.g., microwave spectrometers): calculation of SM by measuring brightness temperature (Tb), soil dielectric properties, and soil temperature.

• Active microwave (e.g., synthetic aperture radar – (SAR)): calculation of SM by estimating backscatter coefficient and dielectric properties.

• Visible: calculation of SM by determining soil albedo index of refraction.

• Thermal infrared: calculation of SM by measuring surface soil temperature.

The use of active microwave retrievals in measuring surface soil moisture is limited due to low temporal frequency and extreme sensitivity to vegetation and surface roughness (Walker et al. 2004). Passive microwave retrieval is the most effective method for soil moisture monitoring on a global scale due to the direct relationship between soil emissivity and soil moisture. These observations are also valid and applicable in all weather and vegetation conditions. Many microwave radiometers such as Special Sensor Microwave/Imager (SSM/I), Advanced Microwave Scanning Radiometer for EOS (AMSR-E), Soil Moisture and Ocean Salinity (SMOS) are used for soil moisture monitoring. However, the main disadvantage of passive microwave retrievals is their low spatial resolution, which is usually less than 25 km. Therefore, soil moisture retrieved with these observations is usually not sufficient for regional studies, and higher-resolution soil moisture monitoring is required (Kim & Hogue 2012). Methods have been developed to cover the reported limitations of microwave spectrum retrievals (low penetration in some weather and vegetation conditions, surface roughness effects, temporal frequency, and coarse spatial resolution) (Merlin et al. 2008; Kim & Hogue 2012). These methods generally fall into two categories: (1) Integration of passive and active microwave soil moisture retrievals to improve the coarse spatial resolution of data (Liu et al. 2012) and (2) integration of soil moisture data from microwave remote sensors and optical remote sensing soil moisture related products to achieve soil moisture data with relatively fine resolution (Chauhan et al. 2003; Piles et al. 2011; Choi & Hur 2012; Zhao & Li 2013). Data derived from optical remote sensors can provide detailed contextual information on the land surface. In this case, some studies have been made to improve the spatial and temporal signatures of microwave soil moisture by merging with visible and infrared (IR) products. These approaches have attracted significant attention since they do not require extensive ground-based data or model-based output and are relatively easy to implement.

Besides surface soil moisture data derived from microwave sensors, land surface parameters derived from multispectral optical remote sensing data have been used for estimating surface soil moisture since the early 1980s (Carlson et al. 1981, 1990, 1995; Patel et al. 2009; Rahimzadeh-Bajgiran et al. 2012; Leng et al. 2014). The normalized difference vegetation index (NDVI) is the most generally used vegetation index (VI) for investigating vegetation growth standing and is commonly declared as the greenness index, which indicates vegetation density and chlorophyll content of vegetation instead of the water condition of an area. Therefore, so as to observe water stress, there's a requirement for an additional sensitive indicator to NDVI. The values of land surface temperature (LST) measured high in dry conditions because of the dearth of soil moisture, and LST is used as a proxy for estimating the state of water stress. The NDVI and LST together will give crucial info on the status of vegetation and soil moisture (Gillies et al. 1997; Sandholt et al. 2002; Wan et al. 2004). Following this path, many studies have examined the appearance of a triangular or trapezoidal shape when drawing VI against LST (Carlson 2007). The appearance of the LST/VI triangular space is due to the low variations of LST over vegetated areas, and the high variations of LST in bare land areas (Sandholt et al. 2002). The focus of these studies is to analyze the biophysical characteristics enclosed in the LST/VI scatter space and to establish a relationship between these biophysical characteristics and the estimation of surface soil moisture (Carlson et al. 1990; Gillies et al. 1997; Petropoulos et al. 2009). Combining microwave retrievals and optical spectrum products of soil moisture is the most generally used technique for downscaling the coarse spatial resolution of soil moisture data retrieved from microwave observations. The LST/VI triangle space based models are the earliest and most commonly used methods. Based on the characteristics of the LST/VI feature space, Chauhan et al. (2003) first integrated passive microwave observations of soil moisture and optical remote sensing products by using a downscaling factor created based on the relatively high spatial resolution optical remote sensing data to improve the spatial resolution of soil moisture data. The method was straightforward to be enforced and commonly used since then (Ray et al. 2010; Piles et al. 2011; Choi & Hur 2012). Merlin et al. (2009, 2010) developed a spatial downscaling method by employing a semi-empirical soil evaporative efficiency model and integrating microwave retrievals of soil moisture and optical spectrum products, and this model is biophysically more accurate than the methods based on the LST/VI triangle space. However, because it requires several field measured biophysical parameters, the Merlin method cannot be easily applied. Kim & Hogue (2012) developed the UCLA¹ method, which applied the soil wetness index (Jiang & Islam 2003) obtained from MODIS² images as the downscaling factor to achieve relatively fine resolution soil moisture data from the coarse resolution microwave retrievals of soil moisture. This study approved that the UCLA method had similar results to the Merlin method on downscaling of the coarse spatial resolution surface soil moisture data.

Now, considering the very small number and dispersion of ground stations for measuring soil moisture in Iran and the importance of knowing the spatial and temporal distribution of soil moisture in water resources and agriculture studies, this study aims to estimate the soil surface moisture using remote sensing data in Urmia Lake basin. For this purpose, the data of visible, near-infrared and thermal bands of MODIS sensor will be used. Radar data from AMSR-E, AMSR2³, and ESA-CCI⁴ satellites and LPRM⁵ and GLDAS⁶ models will be compared using the ground-based soil moisture data available in the study area. The results of this study can be applied as a practical method for researchers in other basins and also as an effective tool for policy makers and decision makers in the water resources and agriculture sector within the basin.

Direct measurements of soil moisture at three ground stations in selected study areas were used to validate the remote sensing data and the estimates made in this study. Soil moisture in these stations is recorded at different soil depths and on weekly, daily, and hourly time scales. However, since the data and remote sensing methods used in this study estimate the soil moisture at a depth of 5 cm of the soil, only data recorded at a depth of 5 cm were used for validation. In statistical evaluations, the passing days of Terra satellite (MODIS sensor) were selected as the criterion and other observational data and remote sensing data were selected only for the selected 8-day time series from MODIS sensor images. The specifications of the selected stations and the statistical year of each station are given in Table 1.

The remote sensing data, its products and the earth data models used in this research are given in Table 2. Also in this table, the specifications related to the name of the satellite or radar, sensor and model used in the relevant product, the developer organization, as well as their spatial and temporal accuracy are given. The last column of this table lists the statistical year used for each product. The time series selected for all this data and products is the 8-day time series of MODIS sensor products (MOD09Q1 and MOD11A2) for both 2010 and 2014 from Julian day 65 (corresponding to March 6) to Julian day 265 (corresponding to September 22) and a total of 26 images were selected.

In this study, LST values are extracted from MOD11A2 images (Table 2) in °K.

In which NDVI is the vegetation index (normalized vegetation difference index) and a, b, c, and d are the coefficients of the linear equations of the wet and dry edges of the triangular space, respectively.

In these equations, and were equal to the observed value (from ground stations) and the estimated value (from remote sensing data) for soil surface moisture, respectively, and the symbols ̅ and σ represent the mean value and the standard deviation of the corresponding values, respectively. The statistical indices of MAE, RMSE, CRMSE (to avoid possible biases) and model efficiency are used to measure the difference between the estimated and observational data series. The MBE index shows the average estimation tendency to overestimate or underestimate the observed values. A bias value of zero indicates that the estimation was able to predict the observed values well. Positive and negative values also indicate overestimation and underestimation of the estimation method, respectively. The correlation coefficient is also used to measure the relationship between two data sets.

Table 3 presents the coefficients of the linear equations of the dry and the wet edges as well as the correlation coefficient of each corresponding linear fit. As reported in Yang et al. (2008), the temporal variations of these coefficients do not have a specific order. In most cases, however, the slopes of the dry and wet edges have a similar trend, indicating that this scattering space will always be triangular or trapezoidal. In most cases, the value of the correlation coefficient is close to 0.9, which indicates the high accuracy of linear fitting of dry and wet edges.

Table 4(a) and (b) shows the correlation values of temperature variables and NDVI and TVDI with the results of soil surface moisture obtained from different data and models for both 2010 and 2014 years. In Table 4, the symbols * and ** at the top of each correlation number indicate the existence of a significant correlation at the level of 1 and 5% between the two comparable variables, respectively. Also, the sign of the positive and negative values of these numbers indicates the existence of direct and inverse correlations between the variables, respectively. As can be seen from the values in the table, in 2010, the soil surface moisture results of ESA-CCI and LPRM as expected are inversely and directly correlated with temperature and NDVI at 1%, respectively, and are inversely correlated with TVDI, for ESA-CCI it is significant at the level of 5% and for LPRM it is significant at the level of 1%. AMSRE radar results due to what has already been discussed about the nature of the captured data of this product, are only directly correlated with the LST variable at the 1% level. Also, as discussed in the previous section, the results of ESA-CCI and LPRM in 2010 have a direct correlation at the level of 1%, but compared to the AMSRE results, both products have an inverse correlation. In 2014, the results of ESA-CCI and GLDAS soil surface moisture variables with the LST variable are inversely correlated at the level of 1%, with the NDVI having a direct correlation at the level of 5% and with the TVDI are inversely correlated at the level of 1%. The results of LPRM model in 2014 have a direct and inverse correlation with the temperature variable and the NDVI at the level of 1%, respectively, and with the TVDI have a direct correlation at the level of 5%. These results are in accordance with what is discussed about the spatial distribution of soil surface moisture in the previous section. The results of ESA-CCI and GLDAS products in 2014 have a direct correlation at the level of 1%, but with the LPRM product (which results from AMSRE radar data) there is an inverse correlation.

In general, the results of this statistical evaluation confirm the reasonable and acceptable trend of LPRM and ESA-CCI products in 2010 and ESA-CCI and GLDAS products in 2014 in estimating temporal changes of soil surface moisture at the basin level.

Table 5 shows the results of statistical evaluation of soil surface moisture values estimated from remote sensing data with observational data of Tabriz University station for 2010 and Khosroshah and Miandoab stations for 2014 in the form of introduced indicators. According to the values shown in the table, in 2010, the LPRM product with the lowest values of MBE, MAE, RMSE, CRMSD and the highest values of R and EF has the best results and AMSRE product shows the weakest results. ESA-CCI product ranks second in 2010 with weaker statistical index values than the LPRM product. In 2014, in both Khosroshah and Miandoab stations, ESA-CCI product shows the best results except R-correlation coefficient. For both stations, the LPRM product has the weakest results and GLDAS product with the highest correlation coefficient among the three products, but, with weaker results of other indicators compared to the ESA-CCI product, is in second place in 2014.

Given the importance of spatial and temporal monitoring of soil moisture at the basin scale, in this paper, using MODIS satellite images and LST-VI scattering space method, different remote sensing products of surface soil moisture with larger scales (10–25 km) were downscaled to the scale of MODIS's images (∼1 km) by UCLA downscaling method. For this purpose, LST and NDVI variables were extracted using MODIS images. The study of spatial distribution maps of LST and NDVI showed the logical trend of spatial variation of these variables in the study area according to the land use map. Also, the study of temporal change graphs of LST and NDVI showed that estimating the time trend of variables in different land uses is reasonable and acceptable. Then the LST-VI scattering space was plotted for the studied time series and the equations of dry and wet edges were also extracted. Examination of these results showed that in all cases a triangular (or trapezoidal) space is formed, which indicates a wide range of soil moisture values in the study area. By estimating TVDI, studying the spatial distribution maps of the TVDI and its time change diagrams in different land uses showed that the trend of spatial and temporal changes of this index is logical and confirmed the accuracy of data and methods used in the study area. Finally, using the TVDI and the UCLA downscaling method, surface soil moisture maps were extracted at the MODIS images scale for various remote sensing products in the Urmia Lake basin. The study of these maps for 2010 and 2014 years as well as statistical evaluations on the results showed that in 2010 the LPRM product and in 2014 the ESA-CCI product had the best results in estimating surface soil moisture in the study area using the method used in this study.

Regarding the necessity of high spatial resolution in field-scale studies and, on the other hand, the relatively low spatial resolution of MODIS images and other remotely sensed soil moisture products applied in the present study, it is suggested that with the development of new remote sensing soil moisture products with higher spatial resolutions such as the Soil Moisture Active Passive (SMAP), similar studies can be performed using these products as well as optical satellite images with higher spatial resolutions, such as Landsat, SPOT, and ASTER images on smaller areas. Furthermore, to precisely assess the accuracy of products and methods used for remote sensing of surface soil moisture, similar studies could also be conducted in areas with more dense ground stations.

A.R. and M.R.-S. studied and designed the concept; A.R. and J.C. collected data; A.R., M.R.-S., J.C., and A.A.C. performed the analyses and interpretation of results; A.R., J.C., and A.A.C prepared the draft manuscript. All authors reviewed the results and approved the final version of the manuscript.

The datasets generated during and/or analyzed during the current study is available from the corresponding author on reasonable request.

The authors declare there is no conflict.