This paper proposes a novel approach to improve accuracy of weakly-supervised change detection in Synthetic Aperture Radar (SAR) imagery. The method is based on the idea to use an initial small size labeled dataset for synthetic training data generation (STDG) using an Ensemble of Self-Organizing Maps (ESOM). The resulting synthetic data set substitutes the initial small set of labeled authentic data and it is used to train a simple deep neural network (DNN) classifier. The proposed DNN architecture consists of 6 layers: an input layer, two Long Short-Term Memory (LSTM) layers, and three Fully Connected (FC) layers. The proposed method is evaluated on a TerraSAR-X image acquired in the Fukushima region, Japan, before and after tsunami. We have used only 200 labeled data (100 for the "change" class and 100 for the "no-change" class). As benchmark method, we have considered a Support Vector Machine (SVM) classifier. The experimental results confirmed the effectiveness of the proposed approach.

This paper proposes a new neural approach for supervised change detection in Synthetic Aperture Radar (SAR) images using Training Data Generation based on Concurrent Self-Organizing Maps (TDG-CSOM). The proposed model is based on the idea to substitute the authentically labeled sample set with the virtual training data generated by the CSOM system. The proposed change detection algorithm has the following processing stages: (a) TDG-CSOM; (b) Multi-Layer Perceptron (MLP) classifier training using the virtual dataset obtained by TDG-CSOM; (c) concatenation of the corresponding pixels belonging to the SAR bi-temporal image; (d) MLP classification to obtain the change map. The proposed method is implemented and evaluated using a TerraSAR-X image acquired in the Fukushima region, Japan, before and after tsunami. The experimental results confirm the effectiveness of the proposed approach.

The present paper presents a novel approach for semi-supervised classification of remote sensing imagery using {K-Means+(GMM-EM)} clustering cascade followed by selection of an amount of clustered pixels to be added to the training set according to their GMM responsibilities. The proposed method has the following steps: (a) clustering of the multispectral pixels using the cascade composed by K-means followed by the Gaussian mixture model (GMM) with expectation-maximization (EM); (b) selection for addition to the training set of an amount of clustered pixels according to their GMM responsibilities using one of several chosen rules; (c) train a semi-supervised SVM classifier (S3VM) using both the selected clustered pixels and also the few authentically labeled ones; (d) S3VM classification. The performances of the proposed algorithm are evaluated using a LANDSAT 7 ETM+ dataset. According to the experimental results, the proposed semi-supervised classification method has led to a significant improvement in performance over the supervised SVM.

This paper proposes a Genetic Algorithm (GA) approach to clean a given classifier training set for remote sensing image analysis. Starting from an initial set of training data, the new method called GA-Training Label Purifying (GA-TLP) consists of the significant training sample selection using GAs in order to maximize the classifier accuracy. This means to retain the most informative samples and to remove the uncertain, redundant, and misclassified ones. As a result of the selection process, we can obtain a purified training set. The proposed model is implemented and evaluated using a LANDSAT 7 ETM+ image. The experimental results confirm the effectiveness of the proposed approach.

This paper presents an approach for improving performances of the unsupervised classification (clustering) in remote sensing imagery by proposing a technique to combine the classical techniques of K-means clustering and the Gaussian mixture model (GMM) with expectation-maximization (EM). The proposed model means to apply firstly K-means clustering, and to use the result of this technique to compute means μk, covariance matrices Σkand mixing coefficients πk, considered as initialization parameters for the next stage of GMM-EM. The performances of the proposed algorithm for remote sensing image classification using a LANDSAT 7 ETM+ dataset are evaluated. The proposed combined clustering model K-means+GMM-EM, has led to a significant improvement in performance over any of the two single clustering techniques K-means and GMM-EMM.

This paper proposes a novel unsupervised change detection model at pixel level based on Pulse-Coupled Neural Networks (PCNN) combined with Change Vector Analysis (CVA). The proposed algorithm has the following steps: (a) computation of Difference Image (DI) corresponding to the magnitudes of Spectral Change Vectors (SCV); (b) evaluation of the total number of firings of each PCNN neuron for a given interval, by assuming the DI as the input image for PCNN; (c) two-class clustering of the firing accumulator content for each PCNN neuron to deduce pixel change map. The model is evaluated using a Landsat-5 image set acquired on a Mexico area before and after two wildfires. The results confirm the effectiveness of our approach. The exciting novelty of the proposed model over the classical PCNN change detection techniques consist both of its marriage PCNN-CVA and of that it produces an effective change detection at pixel level instead of change detection at patch level, characterizing the previous PCNN change detection approaches.

The paper presents an original neural network approach for automatic target recognition (ATR) in the synthetic aperture radar (SAR) imagery using a pulse-coupled neural network (PCNN) segmentation module combined with a classifier based on virtual training data generation (VTDG) using concurrent self-organization maps (CSOM). The proposed ATR algorithm has the following stages: (a) object detection using PCNN image segmentation; (b) feature selection using Gabor filtering (GF) cascaded with principal component analysis (PCA); (c) support vector machine (SVM) classification using VTDG-CSOM to improve the classifier performances. The proposed model has been applied for the recognition of three classes of military ground vehicles represented by the set of 2987 images of the MSTAR public release database. The experimental results have confirmed the method effectiveness, leading to a total success rate of 97.36%.

This paper presents an original approach for improving performances of the supervised classifiers in remote sensing imagery by proposing a technique to refine a given training set using Ant Colony Optimization (ACO). The new method called ACO-Training Label Cleaning (ACO-TLC) applies ACO model for selection of the significant training samples from a given set of labeled vectors in order to optimize the quality of a supervised classifier. This means to retain the most informative samples and to remove the uncertain or misclassified training samples, which lead to classification errors. As a result of the selection process, we can obtain a purified training set. The proposed model is implemented and evaluated using a LANDSAT 7 ETM+ image. The experimental results confirm the effectiveness of the proposed approach.

This paper proposes a novel neural model for unsupervised change detection in time series of multispectral remote sensing imagery using clustering with Self-Organizing Map (SOM) for automatic pseudo-training sample set selection cascaded with Concurrent Self-Organizing Maps (CSOM) classifier. The proposed algorithm has the following steps: (a) computation of difference image (DI) corresponding to the magnitudes of Spectral Change Vectors (SCVs); (b) SOM clustering to automatically deduce the SCV domain quantization parameters defining the pseudo-training sample set regions (changed, unchanged and uncertain); (c) CSOM classification. The model is evaluated using a Landsat-5 image set acquired on a Mexico area before and after two wildfires. As a benchmark, we have considered the classical method of Bayes theory-EM algorithm for selection of pseudo-training sample set combined with a S3VM classifier. The results confirm the effectiveness of our neural approach. Moreover, the exciting advantage of the proposed model over the classical ones is that it does not require any statistical assumptions regarding changed/unchanged SCVs data and it implies a reduced computational effort.

In this paper, we propose a novel model for automatic annotation of high-resolution Earth Observation (EO) overhead imagery, entirely based on neural networks. The model combines the unsupervised pattern recognition of Self-Organizing Map (SOM) with the supervised classifier of Concurrent SOMs (CSOM). The performances of the proposed method is compared with those of the annotation based on classical statistical techniques of Latent Dirichlet Allocation (LDA) and K-Means. The experiments prove the effectiveness of the proposed method.

This paper presents an original approach for improving performances of the supervised classifiers in hyperspectral remote sensing imagery using generation of synthetic samples to optimize the training set. The proposed model called Virtual Sample Generation by Concurrent Self-Organizing Maps (VSG-CSOM) is based on the idea of improving the training set of a supervised classifier by substituting the initial labeled sample set with the ''virtual” samples generated with a system of concurrent SOMs. We have considered a Spatial-Contextual Support Vector Machine (SC-SVM) classifier, taking into consideration both intraband and also interband pixel correlation. The proposed method is implemented and evaluated using two of the most known data sets of hyperspectral images: Indian Pines and Pavia University. The experimental results prove the significant advantage of the new model.

One applies the neural classifier of Concurrent Self-Organizing Maps (CSOM) for change detection in multispectral multi-temporal remote sensing imagery. The present model of change detection has two main processing stages: (a) feature selection using concatenation algorithm (CON); (b) CSOM classifier. CSOM is a supervised neural classifier whose architecture is composed by a collection of small SOM modules, which use a global winner-takes-all strategy. We have compared the performances of CSOM classifier with those of the following benchmark techniques: Nearest Neighbor (NN), Bayes (likelihood classifier), Multilayer Perceptron (MLP), Radial Basis Function neural network (RBF), and Support Vector Machine (SVM). The considered techniques are evaluated using a LANDSAT 7 ETM+ multi-temporal image. One deduces that CSOM leads to best performance of the considered change detection classifiers for independent optimization of any of the two parameters: TSR (Total Success Rate) or MR (Miss Rate).

This paper is dedicated to multispectral facial image recognition, using decision fusion of neural classifiers. The novelty of this paper is that any classifier is based on the model of Concurrent Self-Organizing Maps (CSOM), previously proposed by first author of this paper. Our main achievement is the implementation of a real time CSOM face recognition system using the decision fusion that combines the recognition scores generated from visual channels {(R, G, and B) or Y} with a thermal infrared classifier. As a source of color and infrared images, we used our VICFACE database of 38 subjects. Any picture has 160 times 120 pixels; for each subject there are pictures corresponding to various face expressions and illuminations, in the visual and infrared spectrum. The spectral sensitivity of infrared images corresponds to the long wave range of 7.5 - 13 mum. The very good experimental results are given regarding recognition score.

We deduce an analytical formula for inversion of a complex Van der Monde matrix. It has applications in signal reconstruction, spectral estimation, system identification, as well as in other important signal processing problems. Previously the inversion of such a matrix has been performed by approximations; we further deduce an exact inversion formula.

Proposes the use of Legendre descriptors (LDs) as features for classification of polygonal closed curves. The normalized cumulative angular function of such a curve is expanded in a Legendre polynomial truncated series whose coefficients are used as shape features called the Legendre descriptors (LDs). By considering several examples of polygonal object classification, the computer simulation shows that the LDs lead to significantly better results (increase of interclass distances), by comparison with the classical Fourier descriptors. It seems that the world of Legendre polynomials is more suitable to approximate a polygonal curve than the world of sinusoidal function.>

Proposes the following cascade for intelligent detection of the presence of binary frequency-shift-keying (FSK) signals corrupted by additive white Gaussian noise: (1) discrete Fourier Transform (DFT) for periodogram estimation, computed at the two modulating frequencies; (2) a specific pattern recognition algorithm in the spectral space IR/sup 2/, consisting of one of the following variants: (a) perceptron; (b) fuzzy perceptron; (c) Bayes. The computer simulation results show the significant improvement of the proposed pattern recognition methods by comparison to the classical technique of detection theory by matched filter. The proposed paper tries to build a bridge between the worlds of communications, signal processing and pattern recognition.>

A picture data compression method consisting of a hybrid cascade having four processing stages is presented. The processing stages are: predictive ordering technique (POT), feedback transform coding (FTC), vertical subtraction of quantized coefficients (VSQC), and predictive coding refinements in the signal space consisting of either overshoot suppression (OS) as a first variant or hybrid block truncation coding (HBTC) as a second one. Each of these stages is described, and reconstructed pictures are presented with their coding fidelity performances (mean square quantization error, mean absolute error and signal-to-noise ratio), using as test pictures a portrait and a LANDSAT image. It is shown that good quality images at low bit rates (0.55 to 1.1 bits/pixel) have been obtained.>

A method of efficient image coding called predictive ordering and linear approximation (POLA) is presented. By taking as reference the previously-received scan line, the pixels of the present scan line are placed in order of decreasing amplitudes. The regression line of the ordered pixels is estimated. The linear approximation is refined by computing the one-bit error pattern and by preserving the first two sample moments of the scan line. Simulation results are given showing good image quality at a low bit rate of about 1 bit/pixel.>