Get Complete Project Material File(s) Now! »
SAR images characteristics
Synthetic aperture radar (SAR) satellites collect swaths of side-looking echoes at range resolution and along-track sampling rate to form an image. Range resolution depends on the bandwidth or pulse duration of transmitted signal and is determined by the pulse length (or 1/bandwidth) and the incidence angle: 2 2 = 0τ = 0 (7).
Where τ is the pulse length and equal 1 ( : bandwidth of the radar), 0 is the speed of light. The factor of 2 accounts for the 2-way travel time of the pulse. Ground range resolution is geometrically related to the slant range resolution by = 2 θ0τ with θ is the look angle. Azimuth resolution describes the ability of an imaging radar to separate two closely spaced scatterers in the direction parallel to the motion vector of the sensor.
Real Aperture Radar (RARs) has an azimuth resolution determined by the antenna beamwidth, so that it is proportional to the distance between the radar and the target (slant-range). For RAR, azimuth resolution can be improved only by longer antenna or shorter wavelength. The use of shorter wavelength generally leads to a higher cloud and atmospheric attenuation, reducing the all-weather capability of imaging radars. Synthetic Aperture Radars (SAR) was developed as a means of overcoming the limitations of (RAR). SAR images have several characteristics that make it unique:
• It provides high-resolution two-dimensional images independent from daylight, cloud coverage and weather conditions.
• It is predestined to monitor dynamic processes on the Earth surface in a reliable, continuous and global way with no effects of atmospheric constituents (multi-temporal analysis).
• The amplitude and phase of the backscattered signals are sensitive to dielectric properties (water content, biomass, ice), to surface roughness (ocean wind speed), to target structure and subsurface penetration.
• It provides accurate measurements of distance (e.g. for interferometry).
In space-based remote sensing, the capability to penetrate through precipitation or into a surface layer is increased with longer wavelengths. The shortest wavelengths (Ka, Ku) are strongly attenuated in the lower layers of neutral atmosphere (troposphere). Long wavelengths (P) in turn are subject to strong scattering while passing through the ionosphere (layer F). The intermediate bands (X, C, S, and L) are therefore the most widely used.
Statistical properties of SAR images
A particular effect to be observed in SAR images is the so-called speckle, which is caused by the presence of elementary scatterers with a random distribution within a resolution cell (for example, in a resolution cell of forest land, the scatterers are the leaves, stems, the trunks, objects on the ground etc.). The coherent sum of their amplitudes and phases results in strong fluctuations of the backscattering from resolution cell to resolution cell. Consequently, the intensity and the phase in the final image are no longer deterministic, but follow instead an exponential and uniform distribution, respectively (C. Oliver & S. Quegan, 2004). The total complex reflectivity for each resolution cell is given by: 0, = ∑ √ exp( ) . exp(− 4 ) (9).
where i is the number of elementary scatterers within the resolution cell.
Although it is commonly referred to as noise, speckle cannot be reduced by increasing the transmit signal power, since it has a multiplicative character, i.e., its variance increases with its intensity. To mitigate speckle a technique known as multi-look is utilized, which is basically a non-coherent averaging of the intensity image (John C. Curlander, 1991, C. Oliver & S. Quegan, 2004). The exponential distribution of a single look intensity image is given by: (2 ² 2 ² ) = 1 − (10).
With L is the number of look, the multi-look intensity is given by: = 1 ∑ =1 (11).
The mean value ̅ = ,̅and the variance ( ) = ∑ =1 ∑ =1².
Data Available for the Study
The data used in this study are Sentinel-1 (S1-A and S1-B) operating in Interferometric WideSwath Mode (IW) at level-1 Ground Range Detection (GRD). In this mode, images are provided at 10 m spatial resolution (single look) with a 250 km swath (within 3 sub-swaths) at VV and VH polarizations. Over the swath, the incidence angle ranges from 29.1° to 46°. The data covers the Mekong Delta every 12 days from October 2014 with S1-A, and every 6 days from October 2016 with S1-A and S1-B, with rapid and free of charge data delivery on https://peps.cnes.fr or https://scihub.copernicus.eu/.
The Sentinel-1 images downloaded from the website (Peps.cnes.fr) are at the level 1-A (Level-1 Ground Range Detected (GRD) products consisting of focused SAR data that has been detected, multi-looked and projected to ground range using an Earth ellipsoid model such as WGS84). These data need to be preprocessed and quality assessed before the analysis.
SNAP software was utilized to preprocess the Sentinel-1 images (Sentinels Application Platform, http://step.esa.int/main/download/). The Sentinel-1 images preprocessing comprises the following steps:
• Multi-looking: to reduce the effect of speckle noise, spatial averaging is applied. However, multi-looking also decreases the spatial resolution of SAR images. For that reason, this step is only applied for the data used at the national scale in order to reduce the volume of the data, by averaging a window of 2×2 (20 m space pixels) (for the dataset in the Mekong Delta, multi-looking was ignored to keep original spatial resolution of 10 m).
• Calibration: conversion to the radar backscattering coefficient sigma nought (σ0) from the digital numbers, which follows the procedure specified by the European Space Agency (ESA, 2017).
• Geo-correction: Due to the topographical variations of a scene and the tilt of satellite sensor, distances can be distorted in the SAR images. Terrain Correction is used to compensate for these distortions so that the geometric representation of the image will be corrected.
• Filtering: A multi-temporal filter as described in subsection 3.3.1 was applied to reduce the speckle noise in SAR images and thus increase the original number of looks in the image to a higher ENL (equation 14), without reducing the spatial resolution.
The required ENL can be assessed in order to meet a given probability of error in the rice/non-rice classification problem, as will be described in Section 5.3. This multi-temporal filter has been developed at CESBIO and implemented using Matlab software.
Radar scattering mechanisms of rice fields
The SAR data have proven ability to distinguish rice from other land cover types because of the specific response of the radar backscattering of vegetation with vertical structure over inundated or wet soil. The interaction between a radar electromagnetic wave and vegetation involves mainly three mechanisms: the scattering from the ground attenuated by the vegetation canopy (surface scattering), the volume scattering, and the multiple scattering between the volume and the ground (volume-surface scattering).
The volume-surface scattering term usually brings a negligible contribution compared to the two others in the usual case of vegetation growing over non-flooded soils. However, in the case of flooded fields or fields with wet soil such as rice paddies, this term becomes dominant when the plants develop because of the double-bounce between the water surface and the plant stems, which are the dominant scatterers in the volume. The different backscattering mechanisms are illustrated in Figure 25.
As described in Chapter 2, for traditional cultural practices, the rice fields are covered with a blade of water during most of their growth cycle. For modern alternate wetting and drying (AWD) practices, the fields are inundated only during certain periods, and the soil is wet for the rest of the season. For transplanted rice, radar backscatters from inundated fields before transplanting are low due to specular reflectance from the water surface. However, with the direct sowing, seeds are sown on wet soil and the backscatter at sowing dates has no more low characteristic values. During the growing period from the vegetative stage, to reproductive stage, radar backscatter increases rapidly which is the consequence of a rapid increase in rice plants height and biomass. The following reproductive phase includes the panicle initiation, heading, and flowering processes. During this phase, the plants stop increasing in height, biomass, and the leaves start to wither and die. Ripening is the final stage with its milk, dough, and mature grain processes.
Electromagnetic models have been used to explain the temporal variation of rice backscatter at X band (Le Toan et al., 1989), C band (Le Toan et al., 1997) and L band (Wang et al., 2005). Most studies simulated the backscatter at HH and VV polarization, and the simulation results indicate that 1) the double bounce backscatter is dominant during a large part of the rice cycle, in particular at C and L-band, 2) the strong increasing temporal variation of rice backscatter during the vegetative phase, 3) the large attenuation in VV polarization due to the vertical structure of rice plant, leading to high HH/VV ratio. Those studies have led to the selection of the backscatter temporal change and the polarization ratio as indicators for detection of rice grown area (Bouvet et al., 2009) (Bouvet and Le Toan, 2011).
However, the previous studies had provided simulations for VV, and HH (e.g. to interpret ERS and ENVISAT ASAR C-band data). Sentinel-1 offers a cross-polarised intensity together with a co-polarised intensity in its dual-polarisation products. Theoretical modelling studies are required for a better understanding of the backscatter in HV or VH polarization. In this study, MIPERS (Multistatic Interferometric Polarimetric Electromagnetic model for Remote Sensing) has been used to simulate the HV backscatter. MIPERS developments have been initiated at ONERA (Villard, 2009) during a PhD work and are being pursued at CESBIO. The data used for detailed description of rice plants until heading stage were from Ribbes et al., (1998), and Le Toan et al., (1997) (for logistical reason, it was difficult to conduct detailed measurements of rice plants at different growth stages in the study region).
The model distinguishes the four scattering mechanisms illustrated in Figure 25:
(1) Volume contribution: simple reflection onto volume scatterers (belonging to the vegetation layer).
(2) Double bounce contributions: considering wave-plant-ground or wave-ground-plant interactions. Specular reflections onto the ground surface are accounted for using the modified Fresnel coefficients.
(3) Triple bounce contribution: coupling terms with the ground surface are accounted twice, so that two specular reflections are considered on the ground surface.
(4) Ground direct contribution: simple reflection onto ground scatterers.
The model simulations were used to interpret the VH backscatter of rice fields in the An Giang province measured over a rice field which follows traditional cultivation (long cycle rice, transplanting and continuous flooding).
The simulation 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 26. This result brought a new insight to the earlier knowledge, which often assigned cross-polarisation backscatter to volume scattering. The result also revealed that the double bounce backscatter shows a lower rate in its increase at 55 days after sowing, resulting in a small decrease in VH backscatter at that stage, and this was interpreted as due to higher attenuation at the booting-heading stage. The experimental data show similar trends to the simulations, despite that the input data describing the plant growth were not derived from the description of the plants observed in the experiment. The work needs to be completed with a dedicated campaign measuring geometric and dielectric properties of the components of the rice canopy under study at different dates during the rice season, in order to interpret in details the scattering mechanisms that occur at different growth stage.
Times series analysis at different polarizations
The time series of 126 Sentinel-1 images from 06/10/2014 to 31/03/2018, with a 12-day revisit until 25/09/2016, then a 6-day revisit period afterwards, have been used to analyse the temporal behavior of radar backscatter over rice fields. The images were preprocessed as described in the previous section before being used to extract the radar backscatter coefficients (σ°) of the 60 sampling rice fields.
Figure 27 shows the VH and VV and VH/VV ratio of backscatter coefficients extracted from the 60 sampled fields. For comparison with optical data, instead of Sentinel-2 data which are often affected by cloud cover, the Proba-V NDVI (Normalized Difference Vegetation Index) product has been used. All the NDVI images employed in this work were downloaded from http://www.vito-eodata.be/. Figure 27 contains the NDVI time series from January 2016 to December 2017. The NDVI values are averaged for the rice fields under study from pixels of 100 m × 100 m of Proba-V NDVI (1-2 pixels per sample).
The data time series from 6/10/2014 to 19/11/2017 clearly show characteristic temporal behavior of the backscatter at VV, VH, and VH/VV for each rice season, with a clear similarity between seasons.
Table of contents :
Chapter 1 Introduction
1.1. Importance of rice
1.2. State of the art in the use of remote sensing for rice monitoring
1.3. Research objectives and thesis structure
Chapitre 1 Introduction (français)
1.1. L’importance du riz
1.2. Etat de l’art de l’utilisation de la télédétection pour le suivi du riz
1.3. Objectifs de recherche et structure du manuscrit
Chapter 2 Rice in the world
2.2. Cultural practices
2.3. Rice growth cycle
2.4. Rice productivity
2.5 Global emissions from rice fields
2.6. Summary on Earth Observation requirements for rice monitoring
Chapter 3 Study region and material
3.1. Study region
3.2. Ground data
3.3. SAR data
Chapter 4 Analysis and interpretation
4.1. Ground data analysis
4.2. Radar backscatter analysis & physical interpretation
4.3. Derivation of Indicators for rice mapping and rice monitoring
Chapter 5 Methodology development
5.1. Calculation of classification features
5.2. Seasonal date selection
5.3. The rice/non-rice mapping algorithm
5.4. Estimation of sowing date
5.5. Detection of long/short cycle rice variety
5.6. Detection of rice phenological stage at S1 acquisition
5.7. Estimation of plant height
5.8. Estimation of crop intensity
5.9. Discussion and conclusion
Chapter 6 Mapping products generation, validation and accuracy assessment
6.1. Mapping products generation, validation and accuracy assessment
6.2. Discussion and conclusion
Chapter 7 Use of Sentinel-1 retrieved information in models estimating rice yield and methane emission
7.2. Description of the models
7.3. Rice production estimation using ORYZA2000 model
7.4. Methane emissions estimation using DNDC model
7.5. Discussion and conclusion-way forward
Chapter 8 Conclusion