Transform methods in signal and image processing generally speaking are easy to use and can play a number of useful roles in remote sensing environmental monitoring. Examples are the pollution and forest fire monitoring. Transform methods offer effective procedures to derive the most important information for further processing or human interpretation and to extract important features for pattern classification. Most transform methods are used for image (or signal) enhancement and compression. However other transform methods are available for linear or nonlinear discrimination in the classification problems. In this paper we will examine the major transform methods which are useful for remote sensing especially for environmental monitoring problems. Many challenges to signal processing will be reviewed. Computer results are shown to illustrate some of the methods discussed.

In this paper, a method to generalize previously proposed Chebyshev Kernel function is presented for Support Vector Classification in order to obtain more robust and higher classification accuracy. By introducing the generalized Chebyshev polynomials for vector inputs, we increase the performance of this kernel function. The simulation results show that the proposed generalized Chebyshev Kernel has better performance than the previously proposed kernel for Support Vector Classification. Early simulation results show that the proposed kernel function yields the best classification results for a Breast Cancer dataset.

Time series statistical models such as autoregressive moving average (ARMA) were considered useful in describing the texture and contextual information of an remote sensing image. To simplify the computation, we use a two-dimensional (2-D) autoregressive (AR) model instead. In our previous research, the 2-D univariate time series based imaging model was derived mathematically to extract the features for further terrain segmentations. The effectiveness of the model was demonstrated in region segmentation of a multispectral image of the Lake Mulargias region in Italy. Due to the nature of remote sensing images such as SAR (synthetic aperture radar) and TM (Thermal Mapper) which are mostly in multi-spectral image stack format, a 2-D Multivariate Vector AR (ARV) time series model with pixel vectors of multiple elements (i.e. 15 elements in the case of TM+SAR remote sensing) are examined. The 2-D system parameter matrix and white noise error covariance matrix are estimated for further classifications. To compute the time series ARV system parameter matrix and estimate the error covariance matrix efficiently, a new method based on modern numerical analysis is developed by introducing the Schur complement matrix, the QR (orthogonal, upper triangular) matrix and the Cholesky factorizations in the ARV model formulation. As for pixel classification, the powerful Support Vector Machine (SVM) kernel based learning machine is applied in conjunction with the 2-D time series ARV model. The SVM is particularly suitable for the high dimensional vector measurement as the "curse of dimensionality" problem is avoided.

A method of modeling an image based on 2-D Time Series that merges with the popular MultiClass Support Vector Machine (SVM) as a generalized linear classifier is proposed. In this paper, we present the classification results on both remote sensing and texture type images. Both Binary SVM Classifier and Multiclass SVM are implemented and discussed. The 2-D Time Series models have been useful in describing the texture and contextual information of an image. The different pattern of classes (segments) may differ significantly in texture and thus time series models can be useful for image segmentation of the images. The image model in our earlier work of using ARMA model based region growing method for extracting lake region in a remote sensing image[l] and the extension to a general image segmentation procedure using time series based region growing for remote sensing images[2] has been extended to using SVM classifier instead of region growing method. The procedure is applied to a city near lake area as shown in figure 2 with encouraging preliminary results in segmenting bodies of water and city building. Also, the USC (U. Southern California) texture images (brick, grass pigtail ...etc) are used in this new SVM classifier method.

We study the problems of stacking in seismic profiles. ICA is considered first as it has been a well studied subject in recent years. In contrast to PCA, the objective of ICA is to extract components with higher-order statistical independence. However ICA clearly has the disadvantage of not taking nonstationarity into consideration, typically experienced in the seismic data processing problems. In this paper we examine instead the use of factor analysis as an alternative but effective way of enhancing the seismic profiles rather than using the conventional Stacking method. Factor analysis is a way to fit a model to multivariate data to estimate the interdependence. In a factor analysis model, the measured variables depend on a smaller number of unobserved (latent) factors. Because each factor might affect several variables in common, they are known as common factors. Each variable is assumed to be dependent on a linear combination of the common factors, and the coefficients are known as loadings. Each measured variable also includes a component due to independent random variability, known as "specific variance"because it is specific to one variable. Contrary to stacking, Factor analysis takes into consideration the scaling of the latent signal and makes explicit use of the second order statistics, obtaining high signal to noise ratio. Moreover, it is compared with principal component analysis and Independent Analysis in processing the synthetic Marmousi dataset.

In the petroleum industry, stacking, one of the principal steps of conventional seismic signal processing, plays an important role in enhancing events and cancelling random and coherent noises by utilizing the predesigned redundancy in the seismic data. This paper demonstrates that by applying an alternative technique, factor analysis, to the same dataset, better subsurface image of the earth can be obtained. Contrary to stacking, it takes into consideration the scaling of the latent signal and makes explicit use of the second order statistics, obtaining higher signal-to-noise ratio. Moreover, factor analysis is compared with principal component analysis and independent component analysis, which can both be realized by neural networks, in processing the synthetic Marmousi dataset.

Active radar sensing is an important method of obtaining inventory information about remote and cloud-covered areas of the world. However, automatic interpretation of SAR images is often difficult due to speckle noise. Appearing as a random granular pattern, speckles seriously degrade the image quality and affect the task of human interpretation and scene analysis. For this kind of speckle removal problem, one of the difficulties is to overcome the tradeoff between noise reduction and preserving significant image details. A novel theory of SAR image restoration and enhancement with independent component analysis (ICA) is proposed. We assume that the speckle noise in SAR images comes from a different signal source, which accompanies but is independent (their statistical characteristics are not same.) of the "true signal source" (image details). Thus the speckle removal problem can also he described as "signal source separation" problem. Then in order to enhance the "true signal source", we classify the basis images and span them into two different signal subspaces, namely "true signal subspace" and "speckle subspace". Finally we build different nonlinear estimators in each signal subspace to recover the original image. In our experiments, the SAR images consist of nine channels of images. We compare our method with two other well known speckle reduction approaches (Kuan filter and Lee filter). The results show that with our method the speckle noise is efficiently removed while at the same time important details (edges in particular) are retained without introducing artificial structures. We further calculate the ratio of standard deviation to mean (SD/Mean) for each image and use it as a criterion for image quality and find that the improvement with our method is more evident for images with ''high level speckle noise".

Time series models have been very useful in describing the texture and contextual information of an image. In remote sensing image segmentation, different pattern classes (segments) may differ significantly in texture and thus time series models can be useful for image segmentation of remote sensing images. In this paper, our earlier work of using ARMA model based region growing method for extracting lake region in a remote sensing image (Chen and Ho, 2003) is extended to a general image segmentation procedure using time series based region growing for remote sensing images with some algorithm changes in improving the texture classification results as well as computer calculation efficiency. A first-order autoregressive image model and a second-order autoregressive moving-average image model are implemented for comparison. One advantage of region growing is that only a small number of seed pixels, based perhaps on reliable ground truth, are needed to represent each pattern class. This is different from typical supervised classification which requires a large number of training samples. As the regions grow some pixels near the border of two adjacent segments may be assigned to two (or more) different segments. A statistical hypothesis testing is performed on such pixels so that each pixel will be given a unique assignment. The procedure is applied to the LANDSAT 5 data base in the area of Italy's MULARGIAS lake region with encouraging preliminary results in segmenting bodies of water and vegetation. In addition, we have tested with good results on USC natural scene data, such as grass, brick, brickwall, plasticbubble, sand... etc

Hyperspectral image data (HSI), with hundreds of high resolution spectral bands, are usually utilized for background characterization. Background characterization can be performed supervised or unsupervised. No matter what classification method is used, the reduction of HSI data dimensionality is first conducted to allow effective feature extraction for classification. Principle component analysis (PCA) is generally used to de-correlate data and maximize the information content in a reduced number of features. This maximization of information is based on the covariance matrix of different spectral bands. The principle components generally contain the background of observed terrain. Some small target or small edge is possibly smoothed or lost. PCA projects the data onto the principal directions. Only the correlation of different spectral bands is used during PCA analysis. However, the correlation of different pixels in the terrain also can be used to compress and reconstruct the HSI data. In this paper, the Gauss-Markov random field (GMRF), which is assumed to be the model of observed terrain, and maximum a posteriori (MAP) estimation, are used to compress and construct the HSI data with PCA. Also the unsupervised classification is used to evaluate the result of this method.

Recently there has been increased interest in the use of the independent component analysis (ICA) for image analysis. ICA can be considered as one approach to component analysis. Among other approaches, the traditional principal component analysis (PCA) is most popular. The component analysis that extracts the most important components of the data is useful for data mining in remote sensing which normally involves a very large amount of data. While PCA method attempts to decorrelate the components in a vector, ICA methods are to make the components as statistically independent as possible. ICA methods are generally more demanding in computation than PCA. We have developed a joint cumulant ICA (JC-ICA) algorithm which can be implemented efficiently by a neural network. As such it is a very useful tool for data mining in remote sensing. The use of the algorithm especially in hyperspectral image analysis will be presented in this paper.

This article presents a new algorithm for blind deconvolution and consequently enhancement for ultrasonic C-scan images. A modified two-dimensional (2-D) H/sub /spl infin// state estimation technique has been utilized that deconvolves the effect of unknown blur in the image and enhances the important edges. A new state-space model, based on redundant representation of constrained 2-D convolution sum, has been developed for the image so that it can be utilized in the H/sub /spl infin// state estimation framework. Satisfactory results have been obtained for real images. The same procedure can also be applied to any other image where localized enhancement can be helpful.

A new approach is presented to use higher order statistics (HOS) to deconvolve the effects of blurring in ultrasonic C-scans. The quantization effects of the mechanical scan-grid and the conversion of the 1D A-scan into one pixel on the image cause blurring. The proposed approach is completely blind to the source or the type of distortion and the formulation is purely two-dimensional. When the blurring function is modeled as an AR process, the image is restored recursively with the application of the inverse filter based on the AR estimate. A significant improvement in the image quality has been demonstrated. In particular, the edges are detected more prominently than in the original image.

In ultrasonic nondestructive evaluation (NDE) of materials, pulse echo measurements are masked by the characteristics of the measuring instruments, the propagation paths taken by the ultrasonic pulses, and are corrupted by additive noise. Deconvolution operation seeks to undo these masking effects and extract the defect impulse response which is essential for identification. We show that the use of higher-order statistics (HOS)-based deconvolution methods is more suitable to unravel the aforementioned effects. Synthetic and real ultrasonic data obtained from artificial defects is used to show the improved performance of the proposed technique over conventional ones, based on second-order statistics (SOS), commonly used in ultrasonic NDE.

Cranberry plants are very sensitive to weather and other conditions. In this paper, the condition of cranberry growth is analyzed through PCA (principle component analysis) of the minimum cranberry spectral match measurement data. Three neural network models are applied to the one-month ahead prediction. The simulation results show the high performance modeling ability of these neural networks. The reliable prediction provided by the dynamic neural networks will be useful for the farmers to monitor and control the cranberry growth process.

For a number of years fiber content analysis has presented a challenging problem because of the complexity of the contents of textile material. This work is much needed for automated systems which ideally should be able to identify nondestructively the fiber content by using the acquired images. This paper presents the use of neural networks for fiber content analysis. More specifically the discrimination between the cashmere fiber and the wool fiber is considered. Neural networks are particularly suitable for fusing different features to maximize the classification performance which is presented in the paper.

The behavior of a neural network when the number of training samples is small is examined by using a large remote-sensing database. The paper also presents a new way to reduce the size of the training set without significantly decreasing the classification quality. The effectiveness of the proposed algorithm is examined on the class-sensitive neural network (CSNN) which is known to have a superior classification accuracy over the standard backpropagation trained neural network. It is shown that with a combination of the sample set condensation algorithm and the CSNN, the classification performance degrades only slightly even when the number of training samples is quite small.

This paper discusses the significance of image segmentation via the combination of both statistical and nonstatistical methods based on the hierarchical framework of multiresolution wavelet analysis (MWA) and Gaussian Markov random fields (GMRF). Microcalculations and subtle mass regions are segmented via a fuzzy c-means (FCM) algorithm using localized features. For further enhancement, expected maximization and constrained optimization is applied to a Gibbs distribution defined from the FCM clustered image labels under a Bayesian framework. The effectiveness of this novel algorithm has been clearly illustrated by real mammographic images.

Active sonar classification has been a challenging pattern recognition problem for many years mainly due to the complexity of ocean environment. Improvement of sensors and data acquisition can be very costly and can only provide limited improvement in classification. Neural networks are ideally suited to active sonar classification problems with the potential advantages. In the paper, some active sonar data characteristics are presented, and the performances of several feedforward neural networks are evaluated and compared with the traditional nearest neighbor decision rule. It is concluded that the neural networks studied not only can outperform but also are far more robust than the traditional classifiers.>

Develops a modified probabilistic neural network (PNN) to recognize the international standard book number (ISBN). Effort is made for real time implementation of the recognition process. The authors used a simple but effective feature extraction approach, which takes the meshed averages of an isolated character as feature vector. To achieve better performance they modified the original PNN algorithm. Before normalizing the feature vectors onto a unit hyper-sphere, they are mapped into a one more dimension higher hyper-cube than the feature vector. The best result for correct characters classification is 99.62%.>

Several major deconvolution techniques commonly used for seismic applications are studied and adapted for ultrasonic NDE (nondestructive evaluation) applications. Comparisons of the relative merits of these techniques are presented based on a complete set of simulations on some real ultrasonic pulse echoes. Methods that rely largely on a reflection seismic model, such as one-at-a-time L(1) spike extraction and MVD (minimum variance deconvolution), are not suitable for the NDE applications discussed here because they are limited by their underlying model. L(2) and Wiener filtering, on the other hand, do not assume such a model and are, therefore, more flexible and suitable for these applications. The L(2) solutions, however, are often noisy due to numerical ill conditions. This problem is partially solved in Wiener filtering, simply by adding a constant desensitizing factor q. The computational complexities of these Wiener filtering-based techniques are relatively moderate and are, therefore, more suitable for potential real-time implementations.

Although statistical pattern recognition and artificial neural networks were initially developed independently, they are now closely related. Such relationships are examined in detail in this paper. Some experimental comparison results are also presented.

A brief survey of the existing neural network models for signal/image processing and pattern recognition is presented. A comparison of the back-propagation algorithm for multilayer perception and an adaptive sample set construction procedure offered by Nestor's restricted Coulomb energy network is presented. A performance comparison with real data for ultrasonic nondestructive evaluation of materials is presented.>

The application of neural network technology to nondestructive evaluation (NDE) of materials is considered. Based on a limited ultrasonic NDE study using neural networks, it is demonstrated that the neural network technology is feasible for classifying well-defined defects in NDE. In other NDE problems in which image data can be used, neural networks also are likely to perform very well. At this stage of development, however, the conventional pattern recognition techniques cannot be completely replaced by the neural network technology.>

With the growing demand for image processing, efficient image data compression techniques are much needed for image data transmission and storage. Two image data compression techniques based on the Laplacian Pyramids are examined: the Laplacian Pyramid Vector Quantization (LPVQ) and the Laplacian Pyramid Predictive Compression (LPPC). The LPVQ can provide a reduction from 8 bits per pixel (bpp) to 0.45 bpp with negligible degradation in image quality (0.6% rms reconstruction error), as shown in a test image. The LPPC can provide comparable performance with a proper choice of image model. Both techniques thus can be superior to existing transform coding and predictive compression techniques. The problems of additive noise, pre-and post-processing and hardware realization are also considered.

The area of statistical pattern recognition has experienced about three decades of continued progress. In this paper, early development of the area is examined by reviewing the activities of the first decade. Recent progress in major topics of statistical pattern recognition including classification rules, feature extraction, contextual analysis and cluster analysis, are then reviewed. Particular notes are made on the important contributions of the late Prof. Fu, which have profound impact on the development of statistical pattern recognition. The future outlook of the area is indeed bright as many applications making use of statistical pattern recognition are explored to seek more reliable recognition performance. Efforts on efficient hardware implementation and less costly recognition software packages as well as the combined statistical-structural approach are also likely to increase significantly.

In this paper the use of digital pattern recognition in ultrasonic nondestructive evaluation (NDE) of materials is examined. Emphasis is placed on identifying effective features from time and frequency domains, correlation functions and impulse responses to classify aluminium plate specimen into three major defect geometries: flat and angular cuts, and circular holes. A multi-stage classification procedure is developed which can further determine the angles and sizes for defect characterization and classification. The research clearly demonstrates that modern pattern analysis can significantly improve the nondestructive material evaluation capability of the ultrasonic methods.

Underwater transient signal processing and recognition has been a difficult and challenging problem area for some years. This paper provides a careful review of the progress and presents some new research results. A knowledge-based signal interpretation system is also proposed for transient analysis.

Underwater transient signals are caused by mechanically induced or flow-radiated noises. Characterization and recognition study of the under-water transients is fundamentally important in ship silencing and vulnerability assessment. Past work has used mainly speech processing methods to study the transients. In this paper a critical comparison is made between the transients and speech. Emphasis is placed on the use of low order modelling for representation of the transient signal, segmentation or extraction of events, cluster analysis and classification of transient events. For segmentation, an algorithm is presented that employs an entropy distance based on the covariance estimate of the autoregressive model. By narrowing down to event portions of the transients, which has not been done before, it is shown that effective characterization, segmentation and recognition of underwater transients can be achieved with low-order model at very modest computation cost.

All papers presented at the 1982 IEEE Workshop on Applied Time Series Analysis, sponsored by the Adaptive and Learning Systems Technical Committee of the IEEE Systems, Man, and Cybernetics Society, are summarized. The diverse topics in time series examined by the Workshop should be of interest to many readers of the Society's Transactions.

The paper will review the theory and present in detail the properties, implementation and experimental results of the nonlinear maximum entropy spectral analysis method proposed by P.F. Fougere. Both real and complex signals as well as the two-dimensional (separable) signals are considered. Application to signal recognition is also examined.

This correspondence is concerned with adaptive digital processing to extract impulse-like signal features from the correlated background noise for detection of intruders with the seismic sensor data. Both the adaptive digital filtering and the adaptive Kalman filtering methods are developed and shown to perform nearly the same for a short data segment. For continued processing of a long duration seismic record, the adaptive Kalman filtering considered has better capability to learn the nonstationary data characteristics than the considered adaptive filtering and to adaptively remove the background noise. Detailed experimental results are presented. Other considerations such as the hardware implementation and the relationships among the parameters are also examined.

This paper is concerned with the methodologies in statistical image processing and recognition. Specific areas considered are the following: (1) The decision rules in image recognition and their comparative evaluation under finite sample size condition; (2) Statistical feature extraction techniques for image segmentation with emphasis on the statistical characteristic of textural features; (3) Statistical contextual analysis algorithms for images. Emphasis is placed on the contextual pre processing/postprocessing techniques to implement the optimum decision rules with context; (4) Statistical image modelling techniques including the nonhomogeneous models and the autoregressive models. The software problems involved in these areas are also examined in details.

The first part of the paper deals with unified signal modelling with the teleseismic data by using the linear prediction theory in time series analysis. For an all-pole model, the unique relations among the linear prediction coefficients, cepstral coefficients, reflection coefficients, maximum entropy spectral estimation form a basis for a class of effective feature sets in seismic recognition. A good spectral fit is available with 15 linear prediction coefficients for a 1024- point digital seismic record. The second part of the paper deals with the classification of a seismic data base of 323 records. A short-term spectral feature set provides a 94.17% correct recognition based on 50 good learning samples per class and a weighted - distance one - nearest - neighbor decision rule. This result is superior to the 87.92% correct recognition by using Bayes decision rule with multivariate Gaussian density assumption, and to the 89.32% correct recognition by using autocovariance features as reported by the author at the 1977 IEEE ASSP Conference.

Hartmann [1] and Shehadeh and Chiu [2] have derived expressions for the spectrum of Manchester coded frequency-shift keyed (FSK) signals. The expressions are compared and the conditions under which they coincide are discussed.

The distance measures and the information functions for feature selection are compared. The comparison is based on the available tight upper and lower bounds of the probability of misrecognition, the rates of change of such probability, the effectiveness of a feature subset, and the computational complexity.

Abstract—The continuous output voltage of a radiation sensor is quantized for PCM transmission. A simple technique is presented which enables the computer to continuously detect the presence of a new radiation level by performing data decorrelation, data normalization, and a threshold test. Computer results verify the feasibility of the data processing technique.