Leland E. Pierce

SAR speckle reduction using wavelet denoising and Markov random field modeling

The granular appearance of speckle noise in synthetic aperture radar (SAR) imagery makes it very difficult to visually and automatically interpret SAR data. Therefore, speckle reduction is a prerequisite for many SAR image processing tasks. In this paper, we develop a speckle reduction algorithm by fusing the wavelet Bayesian denoising technique with Markov-random-field-based image regularization. Wavelet coefficients are modeled independently and identically by a two-state Gaussian mixture model, while their spatial dependence is characterized by a Markov random field imposed on the hidden state of Gaussian mixtures. The Expectation-Maximization algorithm is used to estimate hyperparameters and specify the mixture model, and the iterated-conditional-modes method is implemented to optimize the state configuration. The noise-free wavelet coefficients are finally estimated by a shrinkage function based on local weighted averaging of the Bayesian estimator. Experimental results show that the proposed method outperforms standard wavelet denoising techniques in terms of the signal-to-noise ratio and the equivalent-number-of-looks measures in most cases. It also achieves better performance than the refined Lee filter.

Estimation of forest biophysical characteristics in Northern Michigan with SIR-C/X-SAR

A three-step process is presented for estimation of forest biophysical properties from orbital polarimetric SAR data. Simple direct retrieval of total aboveground biomass is shown to be ill-posed unless the effects of forest structure are explicitly taken into account. The process first involves classification by (1) using SAR data to classify terrain on the basis of structural categories or (2) a priori classification of vegetation type on some other basis. Next, polarimetric SAR data at L- and C-bands are used to estimate basal area, height and dry crown biomass for forested areas. The estimation algorithms are empirically determined and are specific to each structural class. The last step uses a simple biophysical model to combine the estimates of basal area and height with ancillary information on trunk taper factor and wood density to estimate trunk biomass. Total biomass is estimated as the sum of crown and trunk biomass. The methodology is tested using SIR-C data obtained from the Raco Supersite in Northern Michigan on Apr. 15, 1994. This site is located at the ecotone between the boreal forest and northern temperate forests, and includes forest communities common to both. The results show that for the forest communities examined, biophysical attributes can be estimated with relatively small rms errors: (1) height (0-23 m) with rms error of 2.4 m, (2) basal area (0-72 m/sup 2//ha) with rms error of 3.5 m/sup 2//ha, (3) dry trunk biomass (0-19 kg/m/sup 2/) with rms error of 1.1 kg/m/sup 2/, (4) dry crown biomass (0-6 kg/m/sup 2/) with rms error of 0.5 kg/m/sup 2/, and (5) total aboveground biomass (0-25 kg/m/sup 2/) with rms error of 1.4 kg/m/sup 2/. The addition of X-SAR data to SIR-C was found to yield substantial further improvement in estimates of crown biomass in particular. However, due to a small sample size resulting from antenna misalignment between SIR-C and X-SAR, the statistical significance of this improvement cannot be reliably established until further data are analyzed. Finally, the results reported are for a small subset of the data acquired by SIR-C/X-SAR. >

Statistical properties of logarithmically transformed speckle

In synthetic aperture radar (SAR) image processing and analysis, the logarithmic transform is often employed to convert the multiplicative speckle model to an additive noise model. However, this nonlinear operation totally changes the statistics of SAR images. In this communication, we first review the statistical properties of speckle noise in both the intensity and the amplitude formats. Then, we derive the probability density functions, the mean values, and the variances to characterize the log-transformed speckle. Finally we discuss the problems introduced by the logarithmic transform on statistical analysis of SAR images. The statistical models developed in this communication will facilitate subsequent SAR image processing tasks based on the additive noise model.

Land-cover classification and estimation of terrain attributes using synthetic aperture radar

Abstract This paper presents progress toward a geophysical and biophysical information processor for synthetic aperture radar (SAR). This processor operates in a sequential fashion to first classify terrain according to structural attributes and then apply class-specific retrievals for geophysical and biophysical properties. Structural and electrical attributes control the radar backscattering from terrain. Experimental data and theoretical results illustrate the sensitivity of synthetic aperture radar to structural properties, such as surface roughness and canopy architecture, to soil moisture and to the aboveground biomass of vegetation and its moisture status. Accurate land-cover classification is of great value in many types of regional- to global-scale modeling, and is also an essential precursor to many techniques for extracting geophysical and biophysical information from SAR data. The sensitivity of SAR to the structural features of terrain leads to landcover classification into simple and easily interpreted structural classes. Knowledge-based, hierarchical classifiers require no a priori information or statistical understanding of a local scene, and are found to yield overall accuracy in excess of 90%. Classification using existing data from the orbital ERS-1 and JERS-1 SARs yield unambiguous land-cover categorizations at greater accuracy and resolution than that afforded by an unsupervised classification of Normalized Difference Vegetation Index as derived from multitemporal AVHRR data. Level I of the SAR terrain classifier differentiates three structural classes; surfaces, short vegetation, and tall vegetation. These classes can be quantized, averaged over the appropriate grid scale and used directly as roughness inputs to general circulation models. Level II of the classifier differentiates vegetation classes on the basis of growth form and leaf type. This level of structural classification is essential in order to improve the performance of semi-empirical approaches for retrieving near-surface soil moisture and aboveground biomass.

Knowledge-based land-cover classification using ERS-1/JERS-1 SAR composites

Land-cover classification of an ERS-1/JERS-1 composite is explored in the context of regional- to global-scale applicability. Each of these orbiting synthetic aperture radars provide somewhat complementary information since data is collected using significantly different frequencies, polarizations, and look angles (ERS-1: C-band, VV polarization, 23/spl deg/; JERS-1: L-band, HH polarization, 35/spl deg/). This results in a classification procedure for the composite image (a co-registered pair from the same season) that is superior to that obtained from either of the two sensors alone. A conceptual model is presented to show how simple structural attributes of terrain surfaces and vegetation cover relate to the data from these two sensors. The conceptual model is knowledge based; and it is supported by both theoretical considerations and experimental observations. The knowledge-based, conceptual model is incorporated into a classifier that uses hierarchical decision rules to differentiate land-cover classes. The land-cover classes are defined on the basis of generalized structural properties of widespread applicability. The classifier operates sequentially and produces two levels of classification. At level-2, terrain is structurally differentiated into man-made features (urban), surfaces, short vegetation, and tall vegetation. At level-2, the tall vegetation class is differentiated on the basis of plant architectural properties of the woody stems and foliage. Growth forms of woody stems include excurrent (i.e., pines), decurrent (i.e., oaks), and columnar (i.e., palm) architecture. Two classes of leaves are considered: broadleaf and needle-leaf. The composite classifier yields overall accuracies in excess of 90% for a test site in northern Michigan located along the southern ecotone of the boreal forest. For the area examined, the SAR-based classification is superior to unsupervised classification of multitemporal AVHRR data supplemented with a priori information on elevation, climate, and ecoregion.

Preliminary analysis of ERS-1 SAR for forest ecosystem studies

The authors examine an image obtained by the C-band VV-polarized ERS-1 SAR with respect to potential land applications. A scene obtained near noon on Aug. 15, 1991, along the US-Canadian border near Sault Ste. Marie is calibrated relative to an array of trihedral corner reflectors and active radar calibrators distributed across the swath. Extensive contemporaneous ground observations of forest stands are used to predict sigma degrees at the time of the SAR overpass using a first-order vector radiative transfer model (MIMICS). These predictions generally agree with the calibrated ERS-1 data to within 1 dB. It is demonstrated that the dynamic range of sigma degrees is sufficient to perform limited discrimination of various forest and grassland communities even for a single-date observation. Furthermore, it is demonstrated that retrieval of near-surface soil moisture is feasible for grass-covered soils when plant biomass is less than 1 tonne/ha. >

Multisensor data fusion using fuzzy concepts: application to land-cover classification using ERS-1/JERS-1 SAR composites

A fuzzy-based multisensor data fusion classifier is developed and applied to land cover classification using ERS 1/JERS-1 SAR composites. This classifier aims at the integration of multisensor and contextual information in a single and a homogeneous framework. Initial fuzzy membership maps (FMMs) to different thematic classes are first calculated using classes and sensors a priori knowledge. These FMMs are then iteratively updated using spatial contextual information. A classification rule is associated to different iterations. This classifier has the following advantages: first, due to the use of fuzzy concepts, it has the flexibility of integrating multisensor/contextual and a priori information. Second, the classification results consist of thematic as well as confidence maps. The confidence map (a classification certainty map representing the degree of certainty in the thematic map) constitutes an important issue in order to evaluate the classification process complexity and the validity of the assumptions. The application of this classifier using ERS-1/JERS-1 SAR composites is shown to be promising.

Validation of the Shuttle Radar Topography Mission height data

The Shuttle Radar Topography Mission (SRTM) provided data for detailed topographical maps of about 80% of the Earths land surface. SRTM consisted of single-pass C- and X-band interferometric synthetic aperture radars (INSARs). In order to utilize SRTM data in remote sensing applications the data must be calibrated and validated. This paper presents The University of Michigans SRTM calibration and validation campaign and our results using recently acquired C-band SRTM data of our calibration sites. An array of calibration targets was deployed with the intention of determining the accuracy of INSAR-derived digital elevation maps. The array spanned one of the X-band swaths and stretched from Toledo, OH to Lansing, MI. Passive and active targets were used. The passive targets included trihedrals and tophats. The locations in latitude, longitude, and elevation of the point targets were determined using differential GPS. We also acquired U.S. Geological Survey (USGS) digital elevation models (DEMs) to use in the calibration and validation work. The SRTM data used in this study are both Principal Investigator Processor (PI) data, which are not the refined final data product, and the ground data processing system (GDPS) data, which are a more refined data product. We report that both datasets for southeastern Michigan exceed the SRTM mission specifications for absolute and relative height errors for our point targets. A more extensive analysis of the SRTM GDPS data indicates that it meets the absolute and relative accuracy requirements even for bare surface areas. In addition, we validate the PI height error files, which are used to provide a statistical characterization of the difference between the SRTM GDPS and USGS DEM heights. The statistical characterization of the GDPS-USGS difference is of interest in forest parameter retrieval algorithms.

Multitemporal Land-Cover Classification Using SIR-C/X-SAR Imagery

Abstract The dual-flight program (April and October) for the SIR-C/X-SAR instrument aboard the shuttle Endeavor was designed expressly to acquire Synthetic Aperture Radar (SAR) imagery at two significantly different seasons. At the Michigan Forests Test Site (MFTS), the April mission occurred at the beginning of the spring thaw and the October mission occurred just prior to and during the fall color change. Four scenes are evaluated at a constant incidence angle. Seven features are extracted from the SAR data for potential use in classification using powers at different frequencies and polarizations. Given multiseason SIR-C/X-SAR imagery, there are three possible approaches in the classifier development: 1) Under the assumption that the scene does not change significantly as a function of time, develop one classification for a set of x scenes using n features, with x times the number of samples per feature; 2) ignore the multiseason availability and develop independent classifications for each scene using n features; 3) develop a true multitemporal classification where N of features equals n (number of features) times x (number of scenes). Each of these is applied using a combination knowledge-based and Bayesian classifier. Level II (roughly forest community) results show that the true multitemporal April/October classification works very well (97%), as do those for the individual scenes (>90%). A pooled classifier works poorly (April=90%, October=77%) and shows that temporal changes in phenology and moisture conditions contribute significant noise in terrain classification.

Toward consistent regional-to-global-scale vegetation characterization using orbital SAR systems

A study was conducted to assess the potential of combined imagery from the existing European and Japanese orbitar synthetic aperture radar (SAR) systems, ERS-1 (C-hand, VV-polarization) and JERS-1 (L-band, HH-palarization), for regional-to-global-scale vegetation classification. For seven test sites from various ecoregions in North and South America, ERS-1/JERS-1 composites were generated using high-resolution digital elevation model (DEM) data for terrain correction of geometric and radiometric distortions. An edge-preserving speckle reduction process was applied to reduce the fading variance and prepare the data for an unsupervised clustering of the two-dimensional (2D) SAR feature space. Signature-based classification of the clusters was performed for all test sites with the same set of radar backscatter signatures, which were measured from well-defined polygons throughout all test sites. While trained on one-half of the polygons, the classification result was tested against the other half of the total sample population. The multisite study was followed by a multitemporal study in one test site, clearly showing the necessity of including multitemporal data beyond a level 1 (woody, herbaceous, mixed) vegetation characterization. Finally, classifications with simulation of backscatter variations shows the dependence of the classification results on calibration accuracy and on naturally occurring backscatter changes of natural surfaces. Overall, it is demonstrated that the combination of existing orbital L- and C-band SAR data is quite powerful for structural vegetation characterization.