Evidence of a topographic signal in surface soil moisture derived from ENVISAT ASAR wide swath data Evidence of a topographic signal in surface soil moisture derived from ENVISAT ASAR wide swath data.

The susceptibility of a catchment to flooding is affected by its soil moisture prior to an extreme 12 rainfall event. While soil moisture is routinely observed by satellite instruments, results from previous work on the assimilation of remotely sensed soil moisture into hydrologic models have been mixed. This may have been due in part to the low spatial resolution of the observations 15 used. In this study, the remote sensing aspects of a project attempting to improve flow 16 predictions from a distributed hydrologic model by assimilating soil moisture measurements are 17 described. Advanced Synthetic Aperture Radar (ASAR) Wide Swath data were used to measure 18 soil moisture as, unlike low resolution microwave data, they have sufficient resolution to allow 19 soil moisture variations due to local topography to be detected, which may help to take into 20 account the spatial heterogeneity of hydrological processes. Surface soil moisture content 21 (SSMC) was measured over the catchments of the Severn and Avon rivers in the South West UK. To reduce the influence of vegetation, measurements were made only over homogeneous pixels of improved grassland determined from a land cover map. Radar backscatter was 1 corrected for terrain variations and normalised to a common incidence angle. SSMC was 2 calculated using change detection. To search for evidence of a topographic signal, the mean SSMC from improved grassland pixels 5 on low slopes near rivers was compared to that on higher slopes. When the mean SSMC on low 6 slopes was 30-90%, the higher slopes were slightly drier than the low slopes. The effect was 7 reversed for lower SSMC values. It was also more pronounced during a drying event. These 8 findings contribute to the scant information in the literature on the use of high resolution SAR soil moisture measurement to improve hydrologic models.


3
Soil moisture can be measured at higher resolution using active SARs rather than passive 13 sensors. Recently there has been increasing interest in estimating soil moisture at local scales 14 using these sensors (Barrett et al., 2009). Two new active SARs suitable for catchment 15 hydrology studies should begin producing data this year. The first of the Sentinel-1 satellites was 16 launched in early 2014. Sentinel-1 is C-band, which will penetrate 1 -2 cm into the soil. 17 Hornacek et al. (2012) have proposed a near real-time automatic system for measuring surface 18 soil moisture at 1km resolution using the Interferometric Wide Swath mode of Sentinel-1. This 19 will measure soil moisture to 6% accuracy, and should be high enough resolution for catchment- 20 scale hydrology studies. When the second satellite of the pair is launched 18 months after the 21 first, they should give near daily coverage over Europe. Also, the Soil Moisture Active Passive 22 sensor (SMAP) is due to be launched early this year (Entekhabi et al., 2010). SMAP is L-band, 6 which will penetrate ~5 cm into the soil (Kerr et al., 2001). It is a combined low resolution 1 radiometer and high resolution SAR, which should give 4% soil moisture accuracy in its 9km 2 resolution product. There is also a radar-only 3km product which will be less accurate. However, 3 possibly this will not be high enough spatial resolution for catchment-scale hydrology studies. a result, we have used ASAR data for this study. ASAR Wide Swath (WS) data were acquired 10 from 2003 -2011, giving a long data record. ASAR is C-band, which penetrates soil to 1 -2 cm. 11 ASAR WS has a spatial resolution of approximately 150m (75m pixel size) and a 400km swath 12 width. VV polarisation images were chosen because of their higher capability of vegetation 13 penetration compared to HH polarisation (Kong and Dorling, 2008). A difficulty with ASAR WS 14 is that the time interval between successive scene acquisitions can be irregular in many areas. For 15 example, in the data set used in this study, there were on average two scenes per month, but in 16 several months there were no useable scenes at all. 17 18 There appears to be scant information in the literature relating to the use of high resolution SAR 19 soil moisture measurement to improve rainfall-runoff estimation. Previous soil moisture studies 20 using high resolution SAR have been aimed mainly at estimating surface soil moisture content  period, it proved possible to map the spatial patterns of soil moisture within a small watershed. 16 This showed the top of the watershed drying out quicker than the floodplain. Roberts and Crane 17 (1997), using ground measurements, also showed that an area on a sloping hillside dried out 18 faster than the valley bottom below. 19 20 The object of this paper is to detect whether a topographic signal can be seen in high resolution 21 remotely sensed soil moisture data. Such a signal may be useful information for a hydrologic 22 model to be able to account for spatial heterogeneity in hydrological processes in relation to 23 8 flood-producing rainfall-runoff events (e.g. Roberts and Crane 1997). The paper is an 1 observational study, and contains no modelling. A subsequent paper will investigate whether the 2 assimilation of these data into a hydrologic model is able to improve runoff prediction. The area considered in this study covered the catchments of the Severn and Avon rivers in the   15 16 The ASAR WS processing chain is shown in fig. 2, and the most important steps are described  2) Incidence angle normalisation. A local incidence angle normalisation is applied for the 13 improved grassland land cover class. containing 25m pixels was averaged to produce 75m pixels to correspond to the ASAR WS pixel 8 size. Because the ASAR WS spatial resolution is twice its pixel size, improved grassland pixels 9 were only selected if a central 75m pixel and its border of 25m pixels were all classed as 10 improved grassland. This avoided edge effects and ensured more homogeneous improved 11 grassland pixels, so that problems caused by mixed pixels could be reduced. Pixels of other land 12 cover types (arable, woodland, urban, water, etc) were ignored. Approximately one-third of the 13 Severn/Avon region is classed as improved grassland, giving a substantial pixel sample size for 14 measurements. 15 16 Radar backscatter generally shows a strong dependence on local incidence angle, with 17 backscatter decreasing strongly with increasing incidence angle over sparsely vegetated terrain    scenes, a difficulty is that wet scenes in which there is open flood water will have very low 1 minimum values at the affected pixels, and these could be misinterpreted as being very dry.

2
Many of these pixels can be removed from consideration by rejecting pixels having backscatters 3 less than a low threshold (-14db), but some mixed pixels covering part land and part flood may 4 remain. The solution that has been adopted is to calculate σ 0 dry at a pixel by taking the average of which are selected as those whose mean backscatter is in the highest quartile of all the scenes.

9
For example, for the improved grassland pixel in fig. 5, σ 0 dry = -13.0db, σ 0 wet = -6.3db, and the 10 sensitivity range is 6.7db.  Here W min was set to the wilting level and W max to the field capacity for clay soil. In order to search for evidence of a topographic signal in the remotely sensed relative SSMC, we 12 compared the mean relative SSMC from improved grassland pixels on low slopes (less than 7%) 13 near rivers (greater than 0.1km and less than 0.8km from a river), with that from improved 14 grassland pixels on higher slopes (7-20%) not necessarily near rivers. Constructing a mean 15 relative SSMC for each class over the whole of the Severn/Avon catchment would not be 16 sensible, as the rainfall history over such a large area would be unlikely to be uniform over the square, the mean relative SSMC of improved grassland pixels in the low slope and higher slope 21 classes was determined. Only 10km squares that contained 700 or more improved grassland pixels in each class were considered. There were about 50 such squares. Observed rainfall data 1 from rain gauges was interpolated over the whole area to each 10km square using block kriging.    Table 1. Regressions of the difference between the mean relative SSMCs on low and higher slopes (low 3 slope meanhigher slope mean) versus the mean relative SSMC on low slopes (a) for a second-order 4 polynomial (y = a + bx + cx 2 ) for all 10km squares for all images, (b) for a linear regression (y = a + bx) 5 for all 10km squares for all images, and (c) for a second-order polynomial for a rainfall scenario in which 6 either <1mm of rain fell on the day of the acquisition and > 3mm fell on the previous day, or <1mm of 7 rain fell on the acquisition day and the previous day and >3mm fell on the day before that. 8 9 From the plot it can be seen that (a) when SSMC L approaches 100%, there is little difference 10 between the mean relative SSMC of the low and higher slopes, (b) when SSMC L is 30 -90%, the 11 higher slopes are slightly drier than the low slopes (from regression R1, a maximum of about 12 1.6% at an SSMC L of 60%), and (c) when SSMC L is low, the low slopes become slightly drier 13 (about 2% maximum) than the higher slopes. 14 15 Fig. 6 has been produced for all 10 km squares with sufficient statistics for all ASAR acquisition 16 dates. It contains samples that occurred during or immediately after rainfall, when it might be 17 expected that the relative SSMC of low slopes near rivers might be the same as that of higher 21 slopes. Fig. 7 shows a similar plot for 10km squares during a drying phase when it is known that 1 rainfall occurred a day or two previously. The samples in the plot are ones for which either (a) 2 <1mm of rain fell on the day of the acquisition and > 3mm fell on the previous day, or (b) <1mm 3 of rain fell on the acquisition day and the previous day and >3mm fell on the day before that. A 4 second-order polynomial has been fitted to the data, with coefficients given in table 1 (identifier 5 R3). The first and second-order coefficients are again significantly non-zero, and the R 2 value of 6 0.212 explains more variance than the polynomial fit for Fig. 6. Fig. 7 shows that, when SSMC L 7 is 35-70%, the higher slopes are drier than the low slopes to a greater extent than in Fig. 6, with 8 the mean relative SSMC difference achieving a maximum of about 2.8% at an SSMC L of 70%.  Figure 7. Mean relative SSMC difference between low slopes near rivers and higher slopes, versus mean 2 relative SSMC of low slopes near rivers, in 10km squares, for a rainfall scenario in which either (a) 3 <1mm of rain fell on the day of the acquisition and > 3mm fell on the previous day, or (b) <1mm of rain 4 fell on the acquisition day and the previous day and >3mm fell on the day before that. The black line is 5 regression R3 of table 1. 6 7 A further factor potentially affecting the difference in mean relative SSMC between low slopes 8 near rivers and higher slopes in a 10km square might be the elevation of the square. Elevations 9 are higher in the west of the region than the east. The influence of elevation was also examined 10 in a regression analysis. Fig. 8 is a plot of the slope s of SSMC D against SSMC L versus the mean 11 elevation of low slopes within a 10km square, for each 10km square. There appears to be no 12 significant correlation between s and the mean elevation of low slopes within a 10km square, for on 45 samples, R 2 = 0.01). While some changes in remotely sensed relative surface soil moisture content have been detected 3 between low slopes near rivers and higher slopes in specific wetness ranges, it is still necessary 4 to show that these are likely to result in corresponding changes in root zone soil moisture RZSM, 5 because it is the latter that will be assimilated into a hydrologic model. The mean relative SSMC from improved grassland pixels on low slopes near rivers was 3 compared to that from similar pixels on higher slopes not necessarily near rivers. The in each of which there were a substantial number of pixels in each class. It was shown that -6 (a) when the mean relative SSMC on low slopes approaches 100%, there is little difference 7 between the mean relative SSMC of the low and higher slopes,

8
(b) when the mean relative SSMC on low slopes is 30 -90%, the higher slopes are slightly drier 9 than the low slopes,

10
(c) when the mean relative SSMC is low, the low slopes become slightly drier than the higher 11 slopes, 12 (d) if a similar comparison was made during a drying phase when rainfall occurred a day or two 13 previously, the higher slopes became drier than the low slopes to a greater extent than in case (b) 14 when the mean relative SSMC on low slopes was 35-70%, 15 (e) there appeared to be no significant correlation between the slope d(SSMC D )/d(SSMC L ) and 16 the mean elevation of low slopes within a 10km square, 17 (f) based on a very limited sample of ground measurements, there appeared to be a linear 18 relationship between remote sensing and ground profile soil moisture measurements. 19 20 This is evidence that a topographic signal can be seen in high resolution remotely sensed surface 21 soil moisture data, which may be useful information for a hydrologic model to be able to account 22 for spatial heterogeneity in hydrological processes. Unfortunately this signal is relatively weak.
However, a further advantage of using ASAR WS data for measuring soil moisture for 1 assimilation into a hydrologic model is their high spatial resolution, which, when combined with 2 a land cover map, allows soil moisture to be measured over single homogeneous pixels. This 3 would not be the case for low resolution microwave sensors, or even for the 1km-resolution soil 4 moisture product from Sentinel-1. While the resolution of Sentinel-1 in Interferometric Wide 5 Swath Mode is higher than that of ASAR WS, the resolution of the latter product has been 6 selected because the averaging of high resolution SAR measurements to a lower spatial  In this study the SSMC from improved grassland pixels on low slopes near rivers was compared 12 to that on higher slopes not necessarily near rivers. This approach was followed because it  Future work will involve the selection of a suitable distributed hydrologic model and