The methods for rice crop mapping and monitoring that are developed and implemented in this project are based on time-series of the standard Sentinel-1 images for land, which are IW (Interferometric Wide-Swath), of 250 km swath width, 5mx20m resolution (1 look), incidence angle in 3 strips from 31° to 46°, 2 polarisations VV and VH, repeat cycle (12 days-and 6 days with Sentinel-1B, after 2016).
The C-band backscatter of rice fields: a state of the art
The backscattering mechanisms
The radar data are well adapted to distinguish rice from other land cover types because of the specific response of the radar backscattering of inundated vegetation. The interaction between a radar electromagnetic wave and vegetation involves mainly three mechanisms: the volume scattering, the scattering from the ground attenuated by the vegetation canopy, and the multiple scattering between the volume and the ground. The last term usually brings a negligible contribution compared to the two others, except in the case of flooded fields or fields with wet soil where it can become dominant as the plants develop, as it has been demonstrated by theoretical models for the case of C-band co-polarized (HH or VV) backscatter at 23° incidence angle , . The different backscattering mechanisms are illustrated in Figure 2.
The co-polarised temporal change method
As a result of the importance of this volume-ground interaction, with the dominant scatterers in the volume being the plant stems, the backscattering coefficients at polarizations HH and VV show a significant increase during the vegetative phase (which lasts 40 to 70 days depending on the rice varieties), right after the low values of the flooding stage, and then decrease slightly during the reproductive phase until harvest. This rice backscatter temporal change was generally observed from ERS, RADARSAT-1 or ASAR to be superior to 8 dB, and sometimes much more , –. This has led to the development of effective rice mapping methods using the temporal change of 23° incidence C-band co-polarized backscatter as a classification feature , , , . This temporal change method was further adapted to time-series of co-polarised (HH) wide-swath ASAR data in the Mekong River Delta . The long revisit (35 days) was tackled by using overlapping acquisitions from adjacent orbits. The variation of incidence angle from the near range to the far range of the image was not found to be problematic as the temporal increase of the backscatter was consistently high over this range of incidence angle. A map of rice cropping intensity has been provided for the whole region, showing strong agreement with statistics of all the provinces in the delta (R2=0.92). With Sentinel-1, the co-polarised temporal change method developed with wide-swath ASAR data can be adapted. Contrarily to ASAR WSM, the overlap between adjacent orbits of S1 is very limited, but this is not necessary given the S1 shorter revisit time (12 days with one satellite, 6 days with two satellites). The influence of the incidence angle however needs to be further investigated.
The co-polarised polarisation ratio method
backscattering: the vertically polarised wave is more attenuated than the horizontally polarized wave, and for that reason the ratio of the HH and VV backscatter intensities is higher than that of most other land cover classes, reaching values around 6-7dB according to a joint analysis of ERS and RADARSAT-1 data ,  and to the modelling of C-band HH and VV , . A rice mapping method using the polarization ratio HH/VV as a classification feature has been developed and validated at the province-scale in Vietnam, using high-resolution (30m) dual-polarization ASAR data . Similar methods have been successfully developed with X-band. Unfortunately, no satellite currently has the ability to provide wide-swath data with dual-polarisation HH and VV capability (Sentinel-1 has only VV&VH or HH&HV), so the co-polarisation ratio method cannot be applied to map rice on larger areas.
The cross-polarised backscatter
Comparatively to co-polarised backscatter, much less effort has been put on the use of cross-polarised backscatter in rice applications. Measurements with a multifrequency polarimetric scatterometer on an experimental paddy field in Japan  have shown that C-band cross-polarised backscatter is correlated to some rice biophysical parameters (Leaf Area Index, canopy height) and has a temporal behaviour similar to that of co-polarised backscatter. However, HV value and its temporal change was thought to be mostly governed by volume scattering, and therefore is not thought to be specific to rice, but also to other crops in the same scene. With the launch of ENVISAT in 2002, it has been the first time that a satellite SAR system could provide cross-polarised C-band data, as part of ASAR dual-polarisation products (HH&HV or VV&VH). Radarsat-2 also offers this possibility since 2007. However, only very few studies using ASAR have been carried out to assess the potential of HV or VH for rice monitoring. Sentinel-1 offers a cross-polarised intensity together with a co-polarised intensity in its dual-polarisation products. The use of the cross-polarised backscatter should therefore be investigated. Theoretical modelling studies are also needed in order to provide a better understanding of the interactions between cross-polarised waves and the rice canopy.
Findings from Georice
Contribution of cross-polarised backscatter
In order to investigate the potential of the cross-polarised backscatter for rice monitoring, we analysed modelling results and compared with VH data from RADARSAT-2 in a preliminary stage.
The modelling results from MIPERS (Multistatic Interferometric Polarimetric EM model for Remote Sensing) are analysed. This electromagnetic model simulates the backscatter of vegetation and was adapted to the description of rice fields. The model showed that at a shallow incidence (40°, which corresponds roughly to the incidence at the center of the Sentinel-1 IW data), the cross-polarised backscatter is also dominated by the double-bounce interaction between the scatterers and the ground, similarly to the co-polarised backscatter, as shown in Figure 3. This result, which is in contradiction with the earlier knowledge, reveals the potential of VH for mapping methods based on the temporal change. The modelling result are compared to C-band images from Radarsat-2, with a configuration similar to that of Sentinel-1 IW. The analysis revealed a good agreement with the MIPERS result, as can be seen in next Figure.
The preliminary analysis of the acquired Sentinel-1 dataset in 2014, 2015 and early 2016 together with the ground data collected in 2015 has led to other findings.
VH backscatter profiles of rice fields vs other targets
The analysis of the VH backscatter profile of rice fields in Sentinel-1 data has provided similar results compared to the Radarsat-2 profiles. Figure 4 shows the backscatter profile of a rice field during two consecutive seasons. This suggests that rice fields can be mapped by using the backscatter temporal change as a classification feature. Figure 4 shows also the backscatter of the main other land use classes present in the Mekong Delta , here, forest, water, and urban areas, which have much more temporal stability, and which have have very different VH backscatter levels. This is a proof of concept for the use of the VH temporal change as a classifier of rice fields.
Effect of incidence angle
The incidence angle (in degrees) of the IWS data used for rice mapping corresponding to the 3 different sub-swaths are as follows:
As the range of incidence angle is quite large (roughly 31° to 45°), it is necessary to assess the variability of the rice backscatter profiles across the incidence range. Figure 5 shows the value of the maximum temporal changeof VV of several rice fields situated in different 2° incidence sub-ranges. The figure shows that the maximum temporal change rises with the incidence, from about 3.2-4.5dB in near range to about 8.5-11.5dB in far range, due to the difference in both the backscatter of smooth surface at the start of the season, and that of the multiple scattering at the peak season. Similar analysis results are obtained with HV, where 3-4 dB was found at near range and 8_9 dB at far range. In all cases however, the maximum temporal increase is high compared to other land use land cover type. The maximum temporal increase of the other targets is found to be around 0.5-1dB at far range. A threshold of 3 dB could be used to map rice at different incidence angles. However, confusion with other targets is expected to be more important at near range. The methodology to calculate the optimal threshold is described in . In this work in the Mekong Delta , the threshold was determined for each of the beams of S1 images. For this purpose, rice fields have been located within each incidence range and their temporal change is analysed to define the relevant threshold. These thresholds can be used for rice mapping in the Mekong Delta using S1 images acquired every 12 days. To extend the method to other region, there is a need to assess experimentally these thresholds, which depend on the backscatter temporal change of other land use land cover types, and also on the time interval between acquisitions (12 days vs 24 days).
Stability across seasons (robustness of algorithm)
In order to assess the robustness of the methods, it is necessary to observe the backscatter profiles over several growing seasons. Since the beginning of the acquisitions of Sentinel-1 data over the Mekong Delta around October 2014, time-series covering almost 5 growing seasons have been acquired. Figure 6, Figure 7 and Figure 8 show the VH, VV and VH/VV backscatter profiles of 9 rice fields situated in the An Giang province. The VH backscatter shows a consistent behaviour at all seasons, with a very homogeneous behaviour across different fields. The VV backscatter behaviour is found to be much more complex and hardly exploitable for the development of robust rice field mapping methods. The VH/VV ratio shows a very interesting behaviour, quite similar to that of VH but with an even smoother behaviour in each season. It is foreseen that the ground contribution is mimised in the ratio, giving rise to higher correlation with the plants indicator and parameters such as NDVI/LAI. This is also observed in other studies (to be published) show that VH/VV is highly correlated to NDVI/LAI for several crop types.Physical interpretation of these temporal variations is on going. These analyses reveal that the VH and VH/VV maximum temporal change arerobust classifier of rice fields, and theycould be used for phenological stage estimation. It is to be noted that in Figs 6-8, the 9 fields have been selected as having the same crop calendar. The curves are much more complicated when fields with shifted calendar are put on the same figure.