This work is part of the preparation for the coupled use of SENTINEL 1 and SENTINEL 2 data (radar and optical sensors, respectively). The aim of this study is to evaluate the potential of XbandSARdatacombined with opticaldata to estimate soilmoistureoverirrigated grassland areas located in southeastern France. An approach based on the inversion of the WCM using multi-layer neural networks (NNs) was developed. This approach relies on four main steps: (1) parameterize the WCM, (2) simulate learning the SAR synthetic dataset, (3) train the neural networks according to three inversion configurations using a part of the synthetic dataset, and finally (4) apply the trained NNs on synthetic and real datasets to validate the inversion approach. In this paper, section 2 presents the study area. Section 3 describes spatial images and the ground-truth measurements performed in situ. Section 4 describes the methodology. The results are shown in section 5. Finally, section 6 presents the principal conclusions.
information (i.e., FAPAR and FCover extracted from optical images) was used (Table 2). This result confirmed that the use of information with low sensitivity to soilmoisture (HV or/and VV) in addition to one polarization with high sensitivity to soilmoisture (HH) does not improve the soilmoisture estimate. The analysis of the error of the mv estimates according to NDVI-values showed that for a NDVI lower than 0.7 (LAI<2.5 m²/m²), the estimate of mv was better than 1.5 vol.% as compared to plots with a NDVI higher than 0.7 (case HH+NDVI : RMSE=4.7 vol.% for NDVI<0.7 against 6.2 vol.% for NDVI>0.7). This result confirmed that the soil contribution in C-band remained important when the vegetation cover was dense (i.e., high values of LAI, BIO, VWC, VEH) and that the soil component could be used to estimate the soilmoisture with acceptable accuracy.
Abstract. The objective of this study is to get a better un- derstanding of radar signal overirrigated wheat fields and to assess the potentialities of radar observations for the moni- toring of soilmoisture. Emphasis is put on the use of high spatial and temporal resolution satellite data (Envisat/ASAR and Formosat-2). Time series of images were collected over the Yaqui irrigated area (Mexico) throughout one agricultural season from December 2007 to May 2008, together with measurements of soil and vegetation characteristics and agri- cultural practices. The comprehensive analysis of these data indicates that the sensitivity of the radar signal to vegetation is masked by the variability of soil conditions. On-going ir- rigated areas can be detected all over the wheat growing sea- son. The empirical algorithm developed for the retrieval of topsoil moisture from Envisat/ASAR images takes advantage of the Formosat-2 instrument capabilities to monitor the sea- sonality of wheat canopies. This monitoring is performed using dense time series of images acquired by Formosat-2 to set up the SAFY vegetation model. Topsoil moisture esti- mates are not reliable at the timing of plant emergence and during plant senescence. Estimates are accurate from tiller- ing to grain filling stages with an absolute error about 9% (0.09 m 3 m −3 , 35% in relative value). This result is attractive since topsoil moisture is estimated at a high spatial resolution (i.e. over subfields of about 5 ha) for a large range of biomass water content (from 5 and 65 t ha −1 ) independently from the viewing angle of ASAR acquisition (incidence angles IS1 to IS6).
Monitoring the spatio-temporal variations in vegetation biophysical parameters and soilmoisture is key information for irrigation and crop management at both the farm level and the irrigation network level. Opticaldata in the visible and infrared spectral range have shown great potential for the mapping and characterization of vegetation biophysical parameters such as the Leaf Area Index (LAI) [5–11], biomass, height, and the Vegetation Water Content (VWC) . Several studies used the Normalized Difference Vegetation Index (NDVI) to estimate the LAI of different crop types (such as wheat, grassland, rice, orchard, corn, and maize) or more complex models based on radiative transfer models combined with neural networks [13–15]. In addition, several studies have used the NDVI to estimate grassland biomass and height [16–19]. Schino et al.  and Payero et al.  compared different vegetation indices over two different sites in central Italy and northwestern USA and found that NDVI provides the most accurate estimation of grass biomass and height. Some studies have used another index known as the Normalized Difference Water Index (NDWI), which is computed using the NIR (near infra-red) and the SWIR (short wave infrared), to estimate vegetation water content [21–25]. Chen et al.  showed that the NDVI and the NDWI allow for similar precision in soybean and corn VWC estimates. Gu et al.  found that the NDWI is more sensitive to grassland drought conditions than the NDVI. The use of the NDVI and the NDWI for estimating vegetation biophysical parameters is limited due to the saturation of values when vegetation is high or very dense with high values of LAI. Payero et al.  reported that the NDVI saturated when the height of alfalfa exceeded 40 cm. Anderson et al.  showed that the NDVI and the NDWI saturate when the LAI of corn and soybean surpassed 3.5 m 2 /m 2 and 4.5 m 2 /m 2 , respectively.
As already mentioned, when registering an optical image to a slant-range SAR image, problems can arise in areas where layover effects are present. This occurs because multiple scatterers located at equal distance to the sensor and different heights appear superimposed at the same image pixel. When using the RPC and a DEM or DSM to find the geographic coordinates of this pixel, several solutions are possible (one for each of these scatterers). While all these solutions are valid and make sense for a SAR image, the co-registered optical image will have some artefacts unless the appropriate solution is chosen. The iterative approach described in the previous subsec- tion will find one of these solutions, which may or may not be the appropriate one. In order to properly handle the ambiguities in layover areas we need to find all of these solutions, and then choose the one that provides a natural looking co-registered optical image.
and slowing along the propagation path. The ionospheric ef- fect is inversely proportional to the square of the radar fre- quency in the microwave part of the spectrum, so that lower frequencies could be more affected by the path delay, par- ticularly near the geomagnetic equator and poles (Masson- net and Feigl, 1998). Conversely, at microwave frequencies the tropospheric unpredictable effect on path delay is essen- tially due to water vapor content and cloud hydrometeors, the former being spectrally non-dispersive, whereas the lat- ter is strongly frequency-dependent (and fairly significant at frequencies above C band). The phase delay in the clear tro- posphere consists of hydrostatic and wet components. Al- though the latter is smaller in magnitude (about 30 cm on av- erage at frequencies less than 10 GHz), it is far more spatially variable than the hydrostatic and ionospheric delays. In fact, changes in the distribution of water vapor are associated with clouds, convection, and storms. In addition, variations result- ing from orographic, frontal, coastal, and seasonal gradients may be present.
depth. Specifically, the MSMS readings on the shallow depth (7 cm below ground) were poorly related with the soilmoisture readings of commercial sensor (R 2 value: 0.1392) ( Figure 2 ), which might be caused by the soil settlement over time after MSMS installation. Soil structure was disturbed during installation and became loose ( Figure 3a ). Although the MSMS kit was inserted firmly into soil on the side wall (as shown in Figure 1a), the contact between MSMS sensor surface and soil would become loose when soil started to settle over time. This impact would be the most significant in the shallow soil, where MSMS sensor was stuck in the slim slot of the rod (as shown in Figure 1a ) and could not move downward along with soil. The shallow soil granules are subject to low pressure and are therefore more vulnerable to easily affected by the external force from the ground such as wind blowing or animal walking ( Figure 3a ). On the contrary, the deep soil handled a higher gravity pressure accumulated from shallow and middle soil layers, which made the structure more condense and the soil granules would not settle easily ( Figure 3a ). In addition, water content (% v/v) and water salinity had been found to vary with soil depth  and water conductivity increased along soil depth due to dissolution of minerals into water , which would also affect the soil resistance  and be reflected by the variation of sensor readings along soil depth.
For VOD computed at VV polarization, results show that the temporal dynamic of the estimated VOD fits generally well with the temporal dynamics of NDVI. In general, a medium to good correlation between VOD and NDVI temporal dynamics was obtained (R 2 about 0.58, 0.39, 0.46, and 0.61 for barley, fallow, oat, and wheat, respectively). However, during the beginning of the senescence period (from 25/04/2018 to 15/05/2018 and from 20/04/2019 to 10/05/2019), VOD and NDVI values became uncorrelated. VOD started decreasing on 25/04/2018 in the first crop growth cycle and on 20/04/2019 in the second crop cycle due to a decrease in the canopy water content while NDVI continued increasing due to the increase in the vegetation photosynthetic activity. Moreover, results showed that in the presence of a high temperature over a well-developed canopy (NDVI reaching its peak value), VOD values computed from the SAR images acquired at ~18h (Ascending mode) were lower than those computed from the SAR images acquired at ~6h (Descending mode). This observation was attributed to changes in the canopy water status which leads to a decrease in VWC during the hot afternoon. Finally, our results showed that the temporal dynamics of VOD values computed from VH polarization do not perfectly match that of NDVI (R 2 lower than 0.4 for barley, fallow, oat, and wheat). It is likely due to the multiple-scattering mechanisms present in VH polarization and which is not considered in WCM formulation. In future, other more complex modelling approaches will be used to better account for these volume scattering effects.
linear. Using either, empirically and physically based soilmoistureretrieval algorithms site specific successful results (R² > 0.9 between observed and modelled soilmoisture) may be obtained for bare soils and vegetated soils at macro- scale (Bindlish et al. 2000, Srivastava et al. 2009). The model performance was generally found to be lower for vegetation covered soils as the backscattering coefficient of the soil is attenuated by the vegetation layer (Ulaby et al. 1982). Some studies found that the correlation between radar backscatter at C-band and soilmoisture was poor at field scale and higher correlation where found at catchment scale where site specific effects seemed to average out (Alvarez-Mozos et al. 2005, Cognard et al. 1995). The vegetation effect on C-band may be even as significant as it is applicable for vegetation biomass retrieval (Mattia et al. 2003, Wigneron et al. 1999). The advantage in using mapping radar techniques compared to microwave radiometers is the higher spatial resolution. Therefore radar data is also proposed to support soilmoisture studies within dis- aggregation procedures using passive L-banddata (Narayan et al. 2006, Piles et al. 2009). The first experiment to estimate soilmoisture from microwave radiometers was performed in the 1970s (Schmugge et al. 1974). L-band penetrates vegetation better than C-band and X-band. Consequently passive microwave remote sensing of soilmoisture at L-band is found as the most promising technique for global soilmoisture monitoring except over dense forests (Prigent et al. 2005, Wagner et al. 2007, Wigneron et al. 2003). The background of this technique is the effect of the dielectric properties of the soil on the natural microwave emission from the soil (Schmugge et al. 1980). The dielectric constant can be calculated as a function of soilmoisture and other soil parameters such as soil texture, soil salinity and bulk density. The most widely used dielectric models within the low frequency range (1-20GHz) are the Dobson Model (Dobson et al. 1985) and the Wang-Schmugge Model (Wang et al. 1980).
distribution function (cdf) matching [16,42–44]. The drivers of the merging approach are data availability and the relative accuracy of the products.
Data availability (at a daily time step) and data sensitivity to vegetation are taken into account to combine the merged active and merged passive datasets. The average vegetation optical depth (VOD) over transitional regions (i.e., regions between sparsely and moderately vegetated areas) is calculated and used as the threshold for separating sparsely- from moderately-vegetated regions outside of the transition zones. Active soilmoisturedata are used for regions with moderate vegetation density (a VOD value higher than the threshold), whereas the passive product is used for (semi-) arid regions (a VOD value lower than the threshold) . When, at a given location in transition zones, the correlation coefficient (R) between the merged active and passive soilmoisture products is greater than 0.65, both products are used . This is done by simply averaging merged passive and merged active products for time steps where both products are available; if only one product type is available, that one is used . In , it has been shown that this procedure increases the number of observations while minimally changing the accuracy of the merged soilmoisture product. Moreover, it has been observed that when the active and passive merged datasets at a given location have an R lower than 0.65, using them both reduces the quality of the merged product relative to the single products.
For the preprocess procedure, original 1-km spatial resolution MODIS are projected to WGS 1984 (17N) in ENVI. The 30-m resolution DEM images are also projected to the same projection system. Resampling procedure is conducted by calculating the mean of all the vegetation index value, LST values and elevation within the larger pixel (40 km ×40 km), the MODIS 1-km resolutiondata is aggregated to 40-km scale. For regression model, three models are utilized by different variables for comparison. The first model takes LST and NDVI image as high-resolution input, the second model prefers LST, NDVI and DEM while the third model process LST, NDVI and TWI data instead. High-resolutionsoilmoisture level is computed based on the regression formula. After the preprocess and model building, SMAP data is downscaled using the regression models stated above. The final step is to validate the downscaling output with the SMAPVEX15 field observation. Results from three regression models are compared with the field data of two different dates. To evaluate the regression model, not only the
which are related to climatic differences. Our results underline the ability of ASAR WS retrieved SM data to track this full spectrum of varying moisture content and seasonal behaviour.
The comparison between soilmoisture time series in each ASAR pixel and in the corresponding ECV cells has shown high R values, larger than 0.56 on average at all sites. This highlights the effectiveness of the ECV SM product in representing the soilmoisture conditions, despite the coarse spatial resolution. A further confirmation of the quality of the ECV SM product is given by the fact that the seasonal R values are quite homogenously distributed over all sites, following the same periodic trend observed by carrying out the regional based analysis (higher R values in spring, lower in winter). By studying the spatial distribution of the correlation between ASAR and ECV SM, it has been observed that the soilmoisture behaviour over those areas characterized by lower altitudes is better described by the ECV SM dataset (high R values). These regions mainly correspond to zones of deep, well-drained, mineral and mineral alluvium soils. On the contrary, areas located at a higher altitude, mainly characterized by poorly-drained soil, show lower correlation between the two satellite SM products. In addition, high slope variability contributes to the further loss of correlation. Indeed, infiltration, drainage and runoff depend on the slope angle. Steeper slopes generally cause lower infiltration rates, rapid subsurface drainage and higher surface runoff . Since ASAR WS acquisitions are not necessarily taken at the same time as the ECV satellite data, different moisture content is likely to be detected by each product over these regions. The lower spatial correlation observed in Pallaskenry, which is characterized by a more complex topography and higher altitude, supports such a hypothesis. 6. Conclusions
moisturedata into the WRF–Noah coupled land– atmosphere model. The background error covariances of the atmospheric control states and land surface soil mois- ture states are estimated separately using the National Meteorological Center (NMC) method (Parrish and Derber 1992). The main objective of this study is to in- vestigate the relative impact of jointly assimilating pre- cipitation and soilmoisturedata on the ability to forecast the two variables as well as atmospheric variables that control the land surface energy balance. We choose to assimilate precipitation and soilmoisture retrievals di- rectly instead of using indirect overland radiance assimi- lation, which is not well understood over frequency channels below 50 GHz. It is important to note that direct assimilation of ground-based precipitation rain rates has been used in the ECMWF operational forecast system (Lopez 2011, 2013). We conduct several numerical ex- periments with the developed assimilation system to as- similate data from the Tropical Rainfall Measuring Mission (TRMM) 3B42 version 7 precipitation (Huffman et al. 2007) using the WRF 4D-Var system and the SMOS soilmoisture retrievals (Kerr et al. 2010) via a WRF–Noah one-dimensional variational data assimilation (1D-Var) system. The results are validated against several reference datasets. The results show that assimilation of both TRMM and SMOS data improves forecast skills of pre- cipitation, soilmoisture, and 2-m air temperature and specific humidity. The validation of 2-day forecasts also shows that the improvement rate due to precipitation data assimilation is nearly constant in time beyond a 6-h win- dow, while the effects of soilmoisturedata assimilation increase throughout the 2-day forecasts.
calculated directly as a soil water mass balance (e.g. Aslyng, 1965; Brereton, Danielov and Scott, 1996; Keane, 2001; Holden and Brereton, 2002). However, various methods are used for the calcula- tion of SMD. For example, two distinct approaches have emerged in Ireland: the model by Brereton et al. (1996) (referred to as the Teagasc model) predicts SMD of well-drained soils, on which any water in excess of field capacity is assumed to be instantly drained, while the model employed by Met Éireann (summarised in Keane, 2001), allows water surpluses to accumulate during wet spells, thus pre- dicting SMD of poorly-drained soils. To date, both approaches have generally adopted SMD parameters that were origi- nally established on a soil in Denmark (Aslyng, 1965).
ing components are calculated using the integral equation model (IEM) developed by Fung and Chen (2004). The snow permittivity is calculated using the strong fluctuation theory (SFT) (Stogryn, 1984). The SFT has been tested and veri- fied in the literature (Wang et al., 2000; Tsang et al., 2007). It is also used in the DMRT model of multilayer snowpack developed by Longepe et al. (2009). This model is capable of simulating the interaction of electromagnetic waves with a layer of snow based on the physical parameters (thickness, optical diameter, snow density). The advantage of this model is the simple implementation and its moderate computation time, which is crucial in order to run the data assimilation process, where the electromagnetic model is repeatedly exe- cuted multiple times. With this model, we can calculate the total backscattering coefficient σ pq 0 for different polarization channels (p, q = H or V) from the physical features of each snow layer, the roughness of air–snow and snow–ice inter- faces, and specific radar illumination (frequency, incidence angle).
1: AGS, United States of America; 2: University of New Hampshire; 3: Dartmouth Hitchcock Medical Center Lake Champlain is a large freshwater lake in the northeast USA that supplies drinking water to hundreds of thousands and supports a diverse socioecological system. Recent research has shown clusters of people diagnosed with Amyotrophic lateral sclerosis (ALS) in the region including adjacent to Bays that undergo chronic cyanobacterial Harmful Algal Blooms (cyanoHABs). ALS is a progressive, fatal disease with an average life expectancy of two to five years from the time of diagnosis. Approximately 90% of cases have no know genetic cause and we hypothesis that a combination of genetic susceptibility and environmental exposure to the cyanotoxin BMAA influence health outcomes. To support our public health investigation we are mapping lake water quality conditions using a multiscale approach. We compiled histroical in situ and concordant Landsat 5 TM and 7 ETM+ and MERIS FR archives to operatiinally map the spatiotemporal dynamics of key lake metrics including chl-a, TP, TN, and PFT. Methods applied an initial data mining routine using exhaustive forward and backward stepwise linear regression using ordinary least squares using an efficient branch-and-bound approach. A correlation matrix was applied to identify potentially correlated or redundant dependent and independent variables. Once these initial data exploration stages were complete, more strategic spatially aware spatiotemporal models were developed. Independent variables (i.e., bands, ratios) were examined for performance and using residuals (i.e., F-stat, adjusted R2, significance values, RMSE, Q-Q plots, Cook’s Distance). Akaike information criterion (AIC) was then applied to a subset of strategic models to further help compare models. Using the multicriteria approach the optimal models selected underwent an n-folds cross validation to provide out of sample evaluation statistics (i.e., goodness of fit, RMSE) for lake attribute predictions. A combination of band ratio and shapefilers were found to have moderate success in mapping lake water quality over time to assess trends and risk hot spots. Simultaneously, during a HAB event in St Labans Bay within Lake CHamplain this past summer, we tasked and analyzed Proba CHRIS, E0-1, Rapid Eye, and Landsat 8 OLI to map the distribution and intensity of characteristics. In situ sampling was conducted to collect and measure metrics and toxin levels in St Albans and Mallets Bays. This includes measurements for microcystin, plankton, anatoxin, among other metrics. We present results from our multiscale assessment and promote remote sensing tools to gain a better understanding of aquatic ecosystem services and smart decision making.
Primarily, there are two approaches to handling biases in data assimilation systems (Dee, 2005): (1) “bias-aware” sys- tems which are built to diagnose and correct the biases in the observations and/or the model forecasts during data assim- ilation integration, and (2) “bias-blind” systems which as- sume the observations and model forecasts to be unbiased. Ideally, biases must be estimated by comparing the observa- tions and/or model states to the true mean states derived, for example, from in situ measurements. However, as noted in Draper et al. (2015), developing spatially distributed bias es- timates is much harder for the land surface, compared to the atmosphere or ocean, since point-scale in situ observations are generally not representative of the spatial scale of re- motely sensed or modeled states, due to the heterogeneity of land. Though there have been a number of studies that rely on online estimation of biases (De Lannoy et al., 2007; Reichle et al., 2010), the common practice in land data assimilation studies is to remove the bias between the observations and the model and to use a bias-blind assimilation approach to cor- rect only short-lived model errors. This is typically achieved by rescaling the observations prior to assimilation, to have the same statistics as the model, using quantile mapping ap- proaches so that the observational climatology matches that of the land model. This approach is easy to implement as a preprocessing step to the data assimilation system and has been used extensively in many land data assimilation studies (Reichle and Koster, 2004; Drusch et al., 2005; Crow et al., 2005; Reichle et al., 2007; Kumar et al., 2009; Liu et al., 2011; Draper et al., 2011, 2012; Hain et al., 2012; Kumar et al., 2012, 2014). A known disadvantage of the approach is that it assumes stationarity in model–observational biases and cannot easily adjust to dynamic changes in bias char- acteristics. Common quantile mapping approaches used for scaling observations into the model’s climatology include the standard normal deviate based scaling (Crow et al., 2005) and the CDF (cumulative distribution function)-matching method (Reichle and Koster (2004); hereafter referred to as RK04). The standard normal-deviate-based scaling matches the first and second moments of the observation and model distri- butions, whereas the CDF matching approach corrects all quantile-dependent biases between the model and observa- tions, regardless of the shape of the distributions.
Lake outlines are produced in the requested projection as shape files from the SAR imagery at 10 m spatial resolution. Water surfaces are very well visible in SAR images because of the generally low backscattering intensity at all microwave wavelengths (Strozzi et al., 2000). We preferred to map the outlines manually because of the challenges and uncertain- ties using automatic methods as a consequence of the speckle of the radar images and of the wind and waves conditions of the lakes, which can increase locally the roughness of the wa- ter surface and thus the backscattering intensity. Supplemen- tary geotiff’s of the geocoded SAR images with a layover and shadow mask are also produced. The resulting informa- tion is thus in a form that can easily be integrated in the end- users’ geographic information system (GIS) and used within the available infrastructure. Consistency tests based, e.g. on multiple independent results from different time periods or sensors, or validation with reference information of indepen- dent origin, are essential to better characterize the error and to estimate the reliability of the product.
der to guarantee the prominent appearance of layover areas in the simulated SAR image. The surface roughness is used to represent direct backscattering but also triple reflections from spatially sep- arated corner reflectors (quasi-direct response from corner tip), which are not represented by LOD-1 and LOD-2 models. The re- flection levels of the simulation are restricted to direct backscat- tering and double reflections as these can be directly related to buildings. Reflection levels beyond double reflection are deac- tivated as they are not representative due to missing details on the facade models (facades represented by plane polygons). The simulated SAR images are generated based on a ray tracing pro- cedure (see details in [Auer, 2011]). Thereafter, the geocoding of the images is conducted using the bounding box information of the scene [Tao et al., 2014] (full scene or individual building). In that regard, differences in the height system are considered for avoiding a shift error in range direction. As an example, the dif- ference between the ellipsoidal height and sea level height of the Munich city center is 45.52 m for the case study presented below.