C.F. Burnette

Bayesian classification of polarimetric SAR images using adaptive a priori probabilities

Abstract Most implementations of Bayesian classification assume fixed a priori probabilities. These implementations can be placed into two general categories: (1) those that assume equal a priori probabilities and (2) those that assume unequal but fixed a priori probabilities. We report here on results of classifying polarimetric SAR images using a scheme in which the classification is done iteratively. The first classification is done assuming fixed (but not necessarily equal) a priori probabilities. The results of this first classification are then used in successive iterations to change the a priori probabilities adaptively. The results show that only a few iterations are necessary to improve the classification accuracy dramatically.

Inference of surface power spectra from inversion of multifrequency polarimetric radar data

During the summer of 1988 an intensive field experiment was conducted in the vicinity of Pisgah lava field in the Mojave Desert of southern California. As part of the experiment, physical properties such as microtopography, composition, soil moisture and dielectric constant at five different sites representing surfaces with r.m.s. heights varying from less than one centimeter to tens of centimeters, were measured. In addition, polarimetric radar images at P-band (68 cm wavelength), L-band (24 cm) and C-band (5.7 cm) were acquired at three different incidence angles with the NASA/JPL airborne imaging radar polarimeter. Using trihedral corner reflectors deployed in the area prior to imaging, the radar images were calibrated to provide σ0 values for each resolution element in each scene. This paper reports on the derivation of the power spectrum of surface microtopography by solution of the small perturbation model for multiple incidence angle and multiple frequency radar data. Power-law fits to the power spectra have exponents (slope in log-log plots) that are nearly the same for all surfaces. These values are close to those from measured microtopography profiles. The offset in log-log plots shows very good correlation with measured power spectrum offsets, however the image-derived offsets are consistently lower than measured values. This may be the result of calibration errors, using the wrong dielectric constants in the inversion, or the fact that not all observed energy was scattered by the surface interface alone.

Data volume reduction for single-look polarimetric imaging radar data

An 8-b quantization scheme to reduce the data volume for single-look complex scattering matrix data measured by polarimetric imaging radar systems is described. The scattering matrices are not symmetrized before compression, thereby retaining information about background noise and system effects. The data volume is reduced by a factor of 3.2. It is shown, with measured data, that the signal to quantization noise ratio for the compression scheme is more than 35 dB for the cross-polarized channels, and more than 45 dB for the copolarized channels. >

Incorporation of polarimetric radar images into multisensor data sets

A technique is presented for registering polarimetric synthetic aperture radar (SAR) data to other aircraft and spaceborne data sets. Resampling is done on the full Stokes matrix, allowing full polarization synthesis on the coregistered data set. Analysis of data acquired over Pisgah Crater in the Mojave Desert, CA, as part of the Mojave Field Experiment shows that the resampling does not seriously affect the pedestal heights of polarization signatures or estimates of RMS heights for smooth to moderately rough surfaces. >

Data Volume Reduction For Single-look Polarimetric Imaging Radar Data

An 8-b quantization scheme to reduce the data volume for single-look complex scattering matrix data measured by polarimetric imaging radar systems is described. The scattering matrices are not symmetrized before compression, thereby retaining information about background noise and system effects. The data volume is reduced by a factor of 3.2. It is shown, with measured data, that the signal to quantization noise ratio for the compression scheme is more than 35 dB for the cross-polarized channels, and more than 45 dB for the copolarized channels.

Classification Of Earth Terrain Cover Using Multifrequency Imaging Radar Polarization Data

There has been considerable interest in applying the additional information that radar polarimetry could provide to classification of earth terrain cover. For example, earlier results of an unsupervised classification algorithm showed that polarization information allows one to classify scattering behavior based on the number of reflections that the wave suffered before returning to the radar receiver. Also, it has been shown that when accurate training data are available, a supervised Bayes classifier yields better performance than the unsupervised classification algorithm and simpler discriminates such as single channel amplitudes, or the phase difference between two channels.