Radar backscatter variability on the field scale
An analysis based on multitemporal Sentinel-1 data

Executive summary

Information on vegetation development and soil moisture on agricultural lands is essential for an effective field management and a variety of other purposes. Sentinel-1 data, as satellite-based, freely available radar data with high temporal and spatial resolution, offer great potential for deriving this information. Known difficulties in the valorisation of the radar data are the complex relationships between biophysical vegetation properties, geophysical soil characteristics and the measured backscatter intensities. A variety of correspondingly sophisticated methods are commonly used to isolate and analyse the parameter of interest. However, complications in utilising the radar data also arise from more basal factors. Speckle, the variation of incidence and look angles, and the existence of artificial objects can distort the temporal course of backscatter intensities at the field level. This study first investigates the reduction of these fundamental interfering influences by a targeted selection of the studied fields as well as appropriate radar data processing. The focus of the work is then on the question to which extent vegetation and soil moisture information can subsequently be derived using simple empirical-statistical methods. With regard to vegetation information, phases of uniform long-term backscatter development are checked for correspondence with phenological stages, and correlations with Leaf Area Index (LAI) are calculated. With respect to soil moisture, linear regression models between short-term soil moisture changes and backscatter changes are built. The results show far-reaching possibilities to infer on vegetation development. For the five-month period of investigation, three to four phases of uniform backscatter development are identified for wheat and maize fields examined in this study. They correspond to certain phenological stages since changes in the structure and water content of the respective plants can essentially be held responsible for the long-term development of the backscatter intensities. Possibilities to derive vegetation information also arise from the strong negative correlations found between LAI and backscatter intensity (R2 > 0.8). At the same time, the decisive influence of vegetation limits the possibilities to derive information on soil moisture changes (R2 < 0.4). Here, additional complications exist, among other things, because of the limited success in correcting angular effects as one of the considered interfering influences. However, the coherence of the results indicates the potential of the applied method, which could be further exploited by expanding the database.

Study area & field data

Radar data & processing

Acquistion settings for used S1 scenes

The table above depicts some figures on the Sentinel-1 data used for the investigation. The products were downloaded in ground range detected (GRD) format and processed locally relying on ESA’s SNAP suite. Precise orbit files were applied to enhance the geolocation accuracy. The backscatter values were radiometrically calibrated to 𝛾0. For geometric correction, range doppler terrain correction was performed using a bicubic-interpolated DTM5. To minimise the temporal angular dependence of the signal, the backscatter values were corrected field by field. The applied correction makes use of the orbit configurations in that all measurements from overpasses with the same relative orbit numbers have the same local incidence angle and radar look direction. Provided that a sufficiently large number of scenes is given, differences in averages of orbitally grouped backscatter values can be attributed to these angular effects. Substracting the orbit-wise average difference to the overall mean backscatter from each single measuerment therefore allows to reduce the angular dependence of the signal. For all further analyses, backscatter values were averaged field-wise considering a buffer area at the edges of the fields.


Despite smaller differences between fields of the same crop species, differences in the temporal development of the backscatter coefficients primarily exist between fields of the different crop species. The similarity of backscatter timeseries across fields of the same crop type gives an first indication regarding the successful exclusion of interfering objects as intended by the selection of the study area. A closer investigation of the distributions of the pixel values within the fields with minor deviations from speckle-dependent distributions expected in case of homogeneous surfaces provides further evidence for this conclusion. Regarding the correction of angle-dependent effects, the regularity of short-term fluctuations of the backscatter values shows that the orbital correction has not eliminated angular effects completely. Nevertheless, a visual comparison to the uncorrected signal as well as a quantitative analyses via autocorrelation plots indicates that reducing originally existing angle-dependent effects was at least partially successful.

In general, maize fields temporary show significantly higher short-term variability of the signal, while a long-term trend is less pronounced than for wheat fields. For maize fields, four phases of signal development can be identified and correlated with the vegetation development.  Analogously, for wheat fields three homogeneous phases of signal evolution can be delineated. For both crop types, the characteristics of the individual backscatter timeseries parts can largely be tracked back to the corresponding phenological development of the plants. The different geometry of wheat and maize plants manifests itself in correspondingly different courses of the backscatter during the same phenological stages.

Relationships between phenological development and backscatter time series for maize (left) and wheat (right)

Regarding soil moisture, correlations between the differentiated time series of in-situ soil moisture values and the differentiated time series of backscatter values is investigated. The applied procedure of differencing, i.e. the subtraction of values between two image dates, forms the foundation of change detection approaches. It assumes surface roughness and vegetation influences to be time-invariant so that backscatter differences can be attributed exclusively to changes in soil moisture. As this condition is met only to a limited extent over the entire investigation period, models are also fitted separately for early stages (March to mid-May) and late stages (mid-May to July). Including the separation by polarisation (VV, VH and the cross-polarisation ratio VH/VV), a set of 18 regression models forms the basis for assessing the possibilities to derive information on soil moisture change.

Positive estimates obtained for the regression coefficients show an increase in backscatter between two scenes of 0.13 dB to 0.47 dB per vol% soil moisture increase, depending on the model. As expected, the backscatter tends to increase with an increase in soil moisture associated with an increase in the dielectric constant. Overall, however, only very weak to weak correlations with maximum R2-values up to 0.25 are observed. The fact that correlations are evident especially for the early period and maize illustrates the importance of vegetation influence. While the lack of vegetation cover is noticeable during the initial phase, the high row spacing as well as the plant structure explain the observation of weakened correlations for maize in the period with denser vegetation cover. In addition to vegetation influences, precipitation and changes in surface roughness may have weakened the relationship between soil moisture and backscatter changes in the present analyses.

Timeseries of backscatter (maize, VV & VH)
compared to precipitation & soil mositure evolution
Correlogram (maize, VV)

Full report