This paper proposes a new approach for high resolution SAR processing which is suitable for real-time processing of highly squinted data. The proposed approach does not use any interpolation and minimizes the storage requirement in the processing. The basic idea is to perform a 1D range and azimuth processing with no range cell migration correction. The defocused impulse response function (IRF) is then correlated with a small 2D kernel (typically 32 X 32 points) which deconvolves the amplitude and phase errors. Several simulation results using the parameters of the E-SAR system of DLR are presented to demonstrate the validity of the new approach.

We present a new 2D squint mode Synthetic Aperture Radar (SAR) processor. It is based on the analytical evaluation of the SAR system transfer function via the stationary phase method. In particular we will underline the space-variance nature of the system function. In our processing approach, this space-variance will be taken into account by evaluating the 2D SAR raw data spectrum on a deformed grid; this is possible by using a non standard Fourier transform. Availability of the 2D spectrum on a deformed grid and of the 'exact' expression of the system transfer function allows the high precision processing of the raw data for squinted geometries.

The data produced by a 2.5 m waveband SAR is processed to two phases. The first phase consists in the preliminary filtering of azimuth counts as the ADC output with their subsequent digital recording on a magnetic carrier. The preliminary processing is required to reduce the input data redundancy, while no data loss occurs and the quality of the resulting image is not degraded. The second phase is carried out on the ground with the use of a general purpose IBM PC system provided with the required software which comprises a meter waveband image synthesis program. The synthesis procedure involves a 'fast convolution' FFT algorithm. In the frequency domain the data is processed with high precision due to the floating point mode of operation. A range migration compensation algorithm forms an intrinsic part of the processing. In the meter waveband the migration value amount to dozens of range pixels at the synthetic aperture edges and the resulting image quality depends on the migration compensation accuracy. A noncoherent storage algorithm has also been developed. The processing is carried out within four partially overlapping beams; after coherent storage in every beam the noncoherent summing up of images obtained at various aspects is performed.

Doppler centroid is one of the most important parameters required by spaceborne SAR signal processing, it must be accurately estimated from echo signals themselves in order to perform a fine motion compensation. So far, the clutter-lock method proposed by Li et al. has been most commonly used to estimate this quantity, however, this method operates well only for homogeneous and quasi-homogeneous sources. In this paper, an optimal amplitude correlation approach is presented to estimate Doppler centroid for spaceborne SAR data from a non-homogeneous scene, it includes the following two steps, (1) coarse estimation: spatial amplitude correlation in range-dimension, (2) fine estimation: image amplitude correlation in range-dimension. The effectiveness of this optimal approach is demonstrated with processing spaceborne SEASAT-A SAR data. Compared with the most commonly used clutter-lock method, the improvement of estimation accuracy of non-homogeneous scene is obtained.

Recently the chirp scaling algorithm has been proposed for high quality SAR processing. The algorithm requires first a transformation of the range uncompressed SAR raw data into the range-Doppler domain, which does not permit conventional techniques for the estimation of the Doppler parameters to be introduced efficiently into the processing. This paper first reviews the so called 'correlation Doppler estimator' (CDE), which was proposed for the estimation of the Doppler centroid in the time domain. This estimation algorithm is further extended in order to allow also the estimation of the Doppler rate. To perform this only bright targets are considered. By continuous calculation of the first coefficient of the auto-correlation function, the slope of the frequency history of the Doppler signal is determined, giving an exact estimate of the Doppler rate. to obtain also an estimation of the Doppler rate for scenes without bright targets but with some contrast, a modified version of the SAC algorithm is presented. Further a new method is presented, which can solve the PRF-ambiguity by means of an evaluation of the signal envelope skew in the range-Doppler domain. This technique does not require any additional range FFT, since it is based on the estimation of the variation of a range centroid as a function of the azimuth frequency. Several results are presented, which show the performance of the proposed new approaches. Some remarks are made regarding the inclusion of the described methods into the extended chirp scaling algorithm.

Traditionally, the production of high quality Synthetic Aperture Radar imagery has been an area where a potential user would have to expend large amounts of money in either the bespoke development of a processing chain dedicated to his requirements or in the purchase of a dedicated hardware platform adapted using accelerator boards and enhanced memory management. Whichever option the user adopted there were limitations based on the desire for a realistic throughput in data load and time. The user had a choice, made early in the purchase, for either a system that adopted innovative algorithmic manipulation, to limit the processing time of the purchase of expensive hardware. The former limits the quality of the product, while the latter excludes the user from any visibility into the processing chain. Clearly there was a need for a SAR processing architecture that gave the user a choice into the methodology to be adopted for a particular processing sequence, allowing him to decide on either a quick (lower quality) product or a detailed slower (high quality) product, without having to change the algorithmic base of his processor or the hardware platform. The European Commission, through the Advanced Techniques unit of the Joint Research Centre (JRC) Institute for Remote Sensing at Ispra in Italy, realizing the limitations on current processing abilities, initiated its own program to build airborne SAR and Electro-Optical (EO) sensor systems. This program is called the European Airborne Remote Sensing Capabilities (EARSEC) program. This paper describes the processing system developed for the airborne SAR sensor system. The paper considers the requirements for the system and the design of the EARSEC Airborne SAR Processing System. It highlights the development of an open SAR processing architecture where users have full access to intermediate products that arise from each of the major processing stages. It also describes the main processing stages in the overall architecture and illustrates the results of each of the key stages in the processor.

The detection of moving targets by a side-looking air- or spaceborne Synthetic Aperture Radar (SAR) is limited due to the bandwidth of the clutter. It has an azimuth spectrum whose bandwidth id determined by the antenna beamwidth in azimuth, the platform speed and the wavelength. Application of the Reflectivity Displacement Method (RDM) allows the detection of moving targets, whose azimuth spectra are within the clutter bandwidth, and also solves the azimuth ambiguity problems of moving targets with high radial velocities. The RDM analyzes the frequency shift between two ground reflectivity functions of adjacent and strongly overlapping azimuth power spectra. It allows the evaluation of the two dimensional Doppler Rate Map from the range compressed radar raw data. By analyzing the Doppler rate map as a function of azimuth and range, one can detect and localize moving targets and evaluate their motion in radial and tangential directions. Experimental results prove the consistency of this method.

This paper gives a comparison of different SAR (Synthetic Aperture Radar) raw data reduction algorithms as applied to E-SAR data (Experimental airborne SAR) and spaceborne ERS1 data. The Block Adaptive Quantizer (BAQ) and a Fuzzy Block Adaptive Quantizer (FBAQ) were selected and analyzed. In addition, different algorithms based on a BAQ, the Fast Fourier Block Adaptive Quantizer (FFT-BAQ) and the Block Adaptive Vector Quantizer (BAVQ) were examined. Signal-to-distortion noise ratios (SDNR) of 11.69 dB (BAQ), 8.00 dB (FBAQ) and 11.94 dB (BAVQ) for E-SAR data and 8.77 dB (BAQ), 5.17 dB (FFT-BAQ) and 9.56 dB (BAVQ) for ERS1 data for a data resolution of 2 bits/sample were achieved with a reduction factor of about 3 for the E-SAR data and 2.5 for ERS1 data.

Tromso Satellite Station (TSS) is the Norwegian national receiving station for ERS-1 SAR data. The TSS Fast Delivery (FD) SAR processor was upgraded during spring 1994 so that the whole processing chain will now be performed in power rather than in voltage. This new FD SAR product from TSS needs absolute calibration, and a calibration constant is therefore estimated. First, a TSS Power-processed ERS-1 SAR image was used. This image covers the ESA transponders in Flevoland. The integration method was used to estimate the backscattered power from the 3 transponders. The Earth ellipsoid, local incidence angle, antenna pattern, range-spread loss, pixel size and RCS were also taken into account in the calculation. The result was a calibration constant of 53.51 dB. An other method is to compare the backscatter from an ESA processed PRI product and a TSS Power-processed FD product acquired at the same place and time. Areas were extracted from two such SAR images, and the pixel values averaged in power. The known calibration constant for the PRI product was used in the comparison of the image products. The calibration constant for the TSS FD product was then found to be 54.20 dB. This is close to the result from using the ESA transponders.

In this paper we discussed problems associated with segmentation based on edge detection by performing a least-squares fit to either the local mean or texture of a SAR image. An important stage in the discussion is the extent to which this algorithm represents an optimum process. We therefore study typical statistical properties of a SAR image of the Amazon rain forest and establish corresponding optimum estimators. We demonstrate that the amplitude is not far from optimum for segmenting the mean by least-squares fitting while both the normalized log of the intensity and the amplitude contrast approximate a maximum likelihood texture measure. We next compare the statistics of these measures with equivalent Gaussians to establish the extent to which a least-squares fit represents the maximum likelihood method for determining edge height and position. Finally theoretical predictions are compared with texture segmentation results on the rain forest example.

In Synthetic Aperture Radar (SAR) and other systems employing coherent illumination to form high-resolution images, the resulting image is generally corrupted by a form of multiplicative noise, known as coherent speckle, with a signal-to-noise ration of unity. This severe form of noise presents singular problems for image processing software of all kinds. This paper describes a segmentation scheme, Merge Using Moments (MUM), for image corrupted by coherent speckle. The image is initially massively over-segmented. A scheme based on examination of the statistical properties (moments) of adjoining regions is employed to improve an over-fine segmentation by merging regions to produce a coarser segmentation. This scheme is employed iteratively until no remaining merge appears valid, at which time a good segmentation is obtained. Segmentation using μm on SAR imagery are given and the results compared to other segmentation schemes. The results of using it on typical SAR images illustrate its potential.

This paper proposes a technique for refining a previous segmentation by using different properties of K-distributed SAR clutter simultaneously. We first consider approximate forms for the K distribution based on a Maximum Entropy approach which assumes that only two moments of the data can be estimated with sufficient accuracy over a small sample region. The choice of moments defines the form of the approximate probability density functions (PDF). After the initial segmentation we then propose a post-processing stage in which the values of the moments of the complete previously-identified segments are assumed exact and an optimum fit to the edge position is defined. We demonstrate that joint estimation of the edge position, based on estimates of the mean of the data and of its logarithm, provides a close approximation to the full K- distribution treatment, while being significantly simpler to implement.

Local second order properties, describing spatial relations between pixels are introduced into the single-point speckle adaptive filtering processes, in order to account for the effects of speckle spatial correlation and to enhance scene textural properties in the restored image. To this end, texture measures originating, first from local grey level co-occurrence matrices (GLCM), and second from the local autocorrelation functions (ACF) are used. Results obtained on 3-look processed ERS-1 FDC and PRI spaceborne images illustrate the performance allowed by the introduction of these texture measures in the structure retaining speckle adaptive filters. The observable texture in remote sensing images is related to the physical spatial resolution of the sensor. For this reason, other spatial speckle decorrelation methods, more simple and easier to implement, for example post-filtering and linear image resampling, are also presented in this paper. In the particular case of spaceborne SAR imagery, all these methods lead to visually similar results. They produce filtered (radar reflectivity) images visually comparable to optical images.

A radar image of the land or sea will contain extended regions of apparently homogeneous textures. However, such regions may contain subtle texture anomalies which indicate the presence of a feature of interest. For this reason, it is important to develop methods of detecting anomalies in textured regions. A strategy for texture anomaly detection is to estimate the multivariate probability density function (PDF) of a texture segment using the whole homogeneous region as training data and then to calculate the likelihood of occurrence of each individual segment within the region given the PDF estimate. Those texture segments with anomalous likelihood values can then be flagged as potential texture anomalies. A crucial element of this strategy is the procedure by which the PDF is estimated.

In this paper we first review an algorithm which performs radar cross- section estimation by using techniques based on simulated annealing. Standard simulated annealing approaches to image restoration attempt to categorize each image element as belonging to one of a small number of predefined image states or values. This is restrictive for tasks such as radar cross-section estimation and we present here an algorithm which is capable of producing a real-valued output. This is achieved by introducing an edge detection stage into the simulated annealing process. This original cross-section estimation is based on a flat region model. This is extended here to include linear sloping regions. Results using this new model are given and a qualitative comparison drawn with the original flat region model approach.

We compare the performance of two segmentation algorithms, previously proven on high resolution airborne synthetic aperture radar (SAR) data, on lower resolution spaceborne SAR data. The algorithms are: RWSEG - iterative multiscale edge detection/segment growing; and ANNEAL - maximum a posteriori radar cross-section reconstruction via simulated annealing. To test the utility and robustness of the algorithms they are applied to ERS-1 PRI images of an agricultural area in the UK and salt playa in Tunisia. These scenes are chosen for the differences between what we would define as useful segmentations of them. A number of tests are applied to segmentation output to measure the homogeneity of segments and the complexity of segment boundaries. The segmentations produced by ANNEAL are generally more detailed than those produced by RWSEG, but take much longer to produce. We also investigate how RWSEG can be used to detect structural change in multitemperal sequences of images. It is found that it is not possible to clearly identify structural similiarites and differences when images are segmented separately and segment boundaries are then overlaid. Much better results occur when a multitemporal sequence is segmented as a single entity.

Measurements made by the AirSAR polarimetric airborne system at longer wavelengths and for some vegetation types are inconsistent with a gaussian model for the data, and are much better described by a multivariate K distribution, although this does not appear adequate in all cases. Parameters suggested by the K distribution model are extracted from scattering matrix data for six different vegetation types at C, L and P bands and for four separate looks. As wavelength increases, spatial and inter-look variability increases markedly. The physical basis of this variability and its implications for data analysis are discussed.

A preliminary study of ERS-1 imagery over a portion of Romanian coastal zone is presented. After compression from 16 bit to 8 bit data and application of adaptive filter to reduce speckle and antenna pattern correction for one of the images, the mean radar return from an offshore region with uniform backscatter has been related to the wind conditions. As a particularity for Black Sea, tidal conditions do not influence much spectral signature of data. Smooth slicks, regions of low backscatter also appear to correlate with coastal outfalls. Meteo and seastate observations have been obtained during ERS-1 data acquisition, which may give an answer after ERS-1 data have been processed in future.

In this paper, we are investigating the optimization of frequencies and polarizations from the viewpoint of target detection in SAR images. Frequency and polarization diversity provided by multifrequency polarimetric SAR opens new ways for more robust detection of ground stationary man-made targets in complex clutter. Due to the scene inhomogeneity, the scattering characteristics of the area within each SAR resolution cell may be nonuniform. Each nonuniform pixel have multiple elementary scattering centers belonging to several classes. Radar returns for each pixel are contributions of the collective interactions of electromagnetric waves scattered by all scattering centers within the pixel. Multifrequency polarimetric measurements for each pixel makes it possible to decorrelate signals from different classes of scatterers within the resolution cell. For the general target-plus-clutter versus clutter case, in this paper is established a general model for target detection in clutter using multi-look multifrequency polarimetric SAR measurements only for the pixel under test. Based upon this model, typical polarimetric covariance matrix parameters at several bands are used to illustrate the procedures to optimize frequencies and polarizations for target detection in clutter.

In this paper we examine the SAR raw signal simulation of extended mountainous natural terrain. In order to cope with this goal we need to consider some problems relative to the evaluation of the backscattering pattern and of the efficient and correct inclusion of the SAR system unit response. In particular, and with regard to the first issue inclusion of the third dimension requires accommodation of its coarse description. Subjective and objective norms in order to judge the simulation results are presented and discussed, together with a number of examples.

An urban scene has very complex variety of length scales ranging from much larger to much smaller tan the wavelength of the radiation emitted by a Synthetic Aperture Radar (SAR). The exact solution to this scattering problem requires the solution of Maxwell's equations for the combination of source and scattering objects present in the scene, which for any reasonable size target area is computationally too intensive to be realistic. Hence while a 'numerically exact' solution at present is not possible, some form of appropriate modeling scheme is used as is usual in electromagnetic problems. The geometrical theory of diffraction (GAD) gives an accurate result with a practical amount of computation. This theory is based on the fact that the most important contributions towards the scattered field come from an area in the neighborhood of some critical points on the scattering surface. For a planar surface, three critical points may be regarded: specular, edge-diffraction and corner-diffraction points. A physical optics version of GAD was taken with the approximate diffraction coefficients derived using physical optics approximations to canonical problems. In this paper, the new model is described in addition to an overview of ray-tracing procedure adopted and its resultant images.

The estimation of the best focusing parameters directly from the data is of particular interest in the processing of SAR imagery whenever accurate orbit and attitude information is not available for the satellite-borne sensor. The basic problem is to determine the doppler rate of chirp signals. Time-frequency distributions have been proposed for such application and they prove accurate enough when echoes from strong point scatterers, present within the images area, are used. However, when natural reflectors instead of artificially deployed corner reflectors are used, results are no longer reliable, usually because of the multiplicity of apparently single scatterers. In these cases simple methods will not work, as their outcomes vary too much because of border effects. Hough transform performs well, but it is computationally heavy if accurate estimates are needed. Here the slope of the principal axis of the time-frequency distribution is used to estimate chirp rate. An iterative procedure is used to limit the influence on the accuracy of the results of border effects, background noise and contributions from smaller scatterers at the same slant range. Results are good. A few percent error in the estimated chirp rates has been observed with ERS-1 data.

We present a study of the region of validity of the first order small perturbation theory applied to rough surface scattering. A detailed quantitative understanding of the range of such validity was obtained by comparison with numerical simulations for the case of a periodic surface, or gratings, varying in one dimension. Scattering of electromagnetic waves from an ensemble of gratings of sufficiently long period will give a good approximation to the case of an infinite rough surface. The comparison procedure undertaken was focussed on the accuracy of the approximation as a function of rms surface height and the incident angles down to a grazing angle. Results for some examples from a range of material properties, from perfect dielectrics to finite conductors, are shown in this paper. Together with a new heuristic also presented, it is observed that for the case of a pure dielectric the first order small perturbation theory for the horizontal polarization remains valid for all incidence angles, while for the vertical polarization it seems to fail if the incidence angle approaches the Brewster angle.

Structured interviews were used to gather information on the commercial requirements and value of SAR data from about 2005 for the global sample of market sectors using cartographic information and spatial statistics. The results show the relative importance of major application areas as commercial prospects. The annual market projection for the period around 2005 using as yet an incomplete global sample is for over 20,000 scenes with a value of the order of $75 million. The most significant markets are expected to be for topographic maps, oceans and ice and agriculture. However, SAR is unlikely to be the sole source of information for any application but will be incorporated with other data as available. The results of the interviews have been used to define customer requirements on the ground segment, which showed that commercial partnership in the mission operations was essential. A hierarchial market model was developed to support the sampling strategy and to relate customer requirements to the SAR design. The work has established a baseline of information on which to plan a commercial SAR mission in the next century.

This paper outlines the principles behind the software architecture design for a real-time processor for airborne SAR data. This processor is implemented on an MIMD parallel computer (the Meiko CS1) using a point-to-point message passing system. The processing algorithms are the result of research by the DRA, Malvern, and are capable of yielding focused, undistorted SAR imagery. Processing functions considered include: initial motion compensation (based on accelerometer data), autofucus with phase correction, and azimuth focusing. Real time processing rates of about 10 MBytes/s are now routinely achieved. We indicate the compromises between processor power, available local memory and communications bandwidth needed to achieve real-time operation. A recent development of the SAR processor has been the addition of a generic post-processing module to allow image interpretation algorithms to operate on the imagery as it is produced. Real-time SAR segmentation has been demonstrated using this facility; ports of other algorithms are planned.

This paper presents the conceptual design, performance analysis and results of the high resolution real-time processor integrated in the experimental airborne SAR-system of the DLR. A Real-Time Subaperture (RTS) approach, which is suitable for airborne real-time SAR processing, was chosen for the hardware implementation. A new approach for high squint SAR processing is adopted, which consists basically of performing a 1D azimuth compression followed by a 2D convolution with a correction function. The main advantages of this approach are the simple hardware implementation, reduced computational and memory requirements and efficient time-domain correction of the range migration. Dedicated on- board hardware, which consists of high speed digital signal processors, programmable logic devices and other integrated circuits, was developed at DLR. The on-board hardware has been successfully tested. The results of the first flight tests in the area of Oberpfaffenhofen (Germany) are shown.

This paper proposed a modified chirp scaling algorithm which is suitable for the processing of highly squinted data with motion errors. By means of azimuth subaperture processing with extended spectral length, the variations of the Doppler centroid in range and azimuth can be accommodated without using block processing with overlap. The new approach, denoted as extended chirp scaling (ECS), is considered to be a generalized algorithm suitable for the high resolution processing of most airborne SAR systems.

This paper extends previous analysis of the errors in the optimum determination of the properties of K-distributed clutter to include the spatial correlation properties. Throughout this analysis of the texture properties of SAR clutter we compare operating in the domain of intensity with using its log or square root. Initially we derive the variance of the correlation coefficients in the three domains and demonstrate the advantage of both the log and amplitude domains over the intensity. In addition we show that a normalized estimator yields a further reduction in the variance. Finally, the results are combined in a derivation of the errors in the correlation length obtained by fitting the correlation coefficients to an assumed exponential autocorrelation function.

An urban scene is a very complex variety of length scales ranging from much bigger to much smaller than the wavelength of the radiation emitted by a Synthetic Aperture Radar (SAR). The exact solution to this scattering problem requires the solution of Maxwell's equations for the combination of source and scattering objects present in the scene, which for any reasonable size target area is computational too large to be realistic. Hence while a 'numerically exact' solution is ruled out but as is usual in electromagnetic problems some form of appropriate modeling scheme is used. In this case we assume that the major contributors to the scattering are the planar surfaces which are generally many times larger than the wavelength of the radiation used. A geometrical optics ray-tracing approach is employed to calculate the incident field on each surface illuminated by the radar system whether by direct or indirect illumination (multiple bounces within the target environment geometry). In this way the large amount of multiple scatter that arises from the combination of dihedral and trihedral corners associated with buildings is taken into account. The strength or amplitude and polarization characteristics of the radiation returning to the radar can be calculated by applying the Rayleigh-Rice scattering theory at each surface.