Classification of high resolution remote sensing data from urban areas is investigated. The main challenge with the classification of high resolution remote sensing image data is that spatial information is extremely important in the classification. Therefore, classification methods for such data need to take that into account. Here, a method based on mathematical morphology is used in order to preprocess the image data. The approach is based on building a morphological profile by a composition of geodesic opening and closing operations of different sizes. In the paper, the classification of is performed on two data sets from urban areas; one panchromatic and one hyperspectral. These data sets have different characteristcs and need different treatments by the morphological approach. The approach can be directly applied on the panchromatic data. However, some feature extraction needs to be done on the hyperspectral data before the approach is applied. Both principal and independent components are considered here for that purpose. A neural network approach is used for the classification of the morphological profiles and its performance in terms of accuracies is compared to the classification of a fuzzy possibilistic approach in the case of the pancrhomatic data and the conventional maximum likelhood method based on the Gaussian assumption in the case of the case of hyperspectral. Different types of feature extraction methods are considered in the classification process.

In this paper we demonstrate how the interaction between innovative methods in the field of computer vision and methods for multi-spectral image classification can help in extracting detailed land-cover / land-use information from Very High Resolution (VHR) satellite imagery. We introduce the novel concept of "geometric activity images", which we define as images encoding the strength of the relationship between a pixel and surrounding features detected through dedicated computer vision methods. These geometric activity images are used as alternatives to more traditional texture images that better describe the geometry of man-made structures and that can be included as additional information in a non-parametric supervised classification framework. We present a number of findings resulting from the integration of geometric activity images and multi-spectral bands in an artificial neural network classification. The geometric activity images we use result from the use of a ridge detector for straight line detection, calculated for different window sizes and for all multi-spectral bands and band-ratio images in a VHR scene. A selection of the most relevant bands to use for classification is carried out using band selection based on a genetic algorithm. Sensitivity analysis is used to assess the importance of each input variable. An application of the proposed methods to part of a Quickbird image taken over the suburban fringe of the city of Ghent (Belgium) shows that we are able to identify roads with much higher accuracy than when using more traditional multi-spectral image classification techniques.

In this paper, we propose a novel scale-driven technique for classification of very high geometrical resolution images. This technique is aimed at obtaining accurate and reliable classification maps by properly preserving the geometrical details present in the images and at the same time by accurately representing the homogeneous areas. The proposed method, on the basis of a multi-scale decomposition of the image under investigation, adaptively selects the proper number of scales to be used in the classification of each single pixel. It is composed of three main steps: i) multiscale/multiresolution decomposition of the considered high resolution image; ii) adaptive selection of the set of best representative scales (levels) of each pixel; iii) classification on the basis of the selected scales. In greater detail, in the first step a multiscale/multiresolution decomposition based on the Gaussian Pyramid analysis is applied to the image under investigation. To correctly classify homogeneous areas while preserving the geometrical information of the scene (details), in the second step we propose to select the set of scales that better represent each specific pixel analyzed. This task is accomplished by a proper adaptive strategy based on the analysis of the neighbor of each pixel at different scales. To obtain the final classification map, in the third step, the selected scales are used as input to a classification architecture composed of different SVM classifiers. Experimental results carried out on a very high geometrical resolution Quickbird image confirm the effectiveness of the proposed technique.

As vegetation time evolution is one of the most relevant information to discriminate the different land cover types, land cover classification requires both temporal and spatial information. Due to the physical properties of remote sensors, this temporal information can only be derived from coarse resolution sensors such as MERIS (300×300 m² pixel size) or SPOT/VGT (1 km² pixel size). In this paper, we propose to use jointly high and coarse spatial resolution to perform an efficient high resolution land cover classification. The method is based on Bayesian theory and on the linear mixture model permitting, through a simulated annealing algorithm, to perform a high resolution classification from a coarse resolution time series.

All the commercial satellites (SPOT, LANDSAT, IRS, IKONOS, Quickbird and Orbview) collect a high spatial resolution panchromatic image and multiple (usually four) multispectral images with significant lower spatial resolution. The PAN images are characterised by a very high spatial information content well-suited for intermediate scale mapping applications and urban analysis. The multispectral images provide the essential spectral information for smaller scale thematic mapping applications such as landuse surveys. Why don't most satellites collect high-resolution MS images directly, to meet this requirement for high-spatial and high-spectral resolutions? There is a limitation to the data volume that a satellite sensor can store on board and then transmit to ground receiving station. Usually the size of the panchromatic image is many times larger than the size of the multispectral images. The size of the panchromatic of Landsat ETM+ is four times greater than the size of a ETM+ multispectral image. The panchromatic image for IKONOS, Quickbird SPOT5 and Orbview is sixteen times larger than the respective multispectral images. As a result if a sensor collected high-resolution multispectral data it could acquire fewer images during every pass. Considering these limitations, it is clear that the most effective solution for providing high-spatial-resolution and high-spectral-resolution remote sensing images is to develop effective image fusion techniques. Image fusion is a technique used to integrate the geometric detail of a high-resolution panchromatic (Pan) image and the color information of a low-resolution multispectral (MS) image to produce a high-resolution MS image. During the last twenty years many methods such as Principal Component Analysis (PCA), Multiplicative Transform, Brovey Transform, IHS Transform have been developed producing good quality fused images. Despite the quite good optical results many research papers have reported the limitations of the above fusion techniques. The most significant problem is color distortion. Another common problem is that the fusion quality often depends upon the operator's fusion experience, and upon the data set being fused. No automatic solution has been achieved to consistently produce high quality fusion for different data sets. More recently new techniques have been proposed such as the Wavelet Transform, the Pansharp Transform and the Modified IHS Transform. Those techniques seem to reduce the color distortion problem and to keep the statistical parameters invariable. In this study we compare the efficiency of eight fusion techniques and more especially the efficiency of Multiplicative Brovey, IHS, Modified IHS, PCA, Pansharp, Wavelet and LMM (Local Mean Matching) fusion techniques for the fusion of Ikonos data. For each merged image we have examined the optical qualitative result and the statistical parameters of the histograms of the various frequency bands, especially the standard deviation All the fusion techniques improve the resolution and the optical result. The Pansharp, the Wavelet and the Modified IHS merging technique do not change at all the statistical parameters of the original images. These merging techniques are proposed if the researcher want to proceed to further processing using for example different vegetation indexes or to perform classification using the spectral signatures.

In this paper, we present two novel methodologies for image-fusion. First we fuse multispectral images from the same satellite (Landsat ETM+) with different spatial resolutions. In this case we show how the proposed method can help improve spatial resolution. In the second case, we fuse multispectral Landsat ETM+ and SAR images combining with a speckle reduction method for the latter. Both algorithms are based on a Gaussian-derivative image transform. This is a multi-channel model for image representation based on the scale-space theory. Moreover, locally rotating the image transform allows better edge reconstruction and restoration from noise. We show that multispectral image fusion with the Hermite Transform preserves the biophysical variable interpretation of the original images.

We present a new hierarchical image segmentation algorithm, which extends a fast multiscale iterated weighted aggregation method originally developed by Brandt et al.(^{1, 2}) to multichannel data. This approach, combining both a region-based and an edge-based approach, results well suited for data (like remote sensing images), which present heterogeneous multiscale characteristics. The generalization of the weighted aggregation method to multispectral remote sensing data requires the choice of a suitable statistical model to describe the multichannel data, together with an appropriate spectral homogeneity criterion. To this purpose, we adopted the multivariate Gaussian model, and the Bhattacharyya distance. For multivariate Gaussian distribution the Bhattacharyya distance results in a closed expression, which is an analytical function of the mean and covariance matrix. Such statistical moments can be computed recursively during the segmentation process, preserving the overall linear complexity of the segmentation process. Our segmentation method has been successfully integrated into Knowledge Information Mining (KIM), a tool developed within an ESA research project for content-based image retrieval from remote sensing large archives. Moreover, the proposed segmentation procedure has been extensively used within the DesertWatch project, an ESA Data User Program (DUP) project for the study and the monitoring of desertification processes from multiple Earth Observation data.

Today's flight safety, especially during aircraft landing approach to an airport, is often affected by adverse weather conditions. One of the promising technologies to increase the pilot's situation awareness is the Enhanced and Synthetic Vision System (ESVS): the combination of sensor vision and synthetic vision systems. In this paper we present one aspect of the sensor vision system, an algorithm that is using only a sequence of infrared images to detect possible runway structures and obstacles on it during the aircraft landing approach. No additional information from database, INS or GPS is used at this processing stage. The algorithm generates several runway and obstacle hypotheses and the final decision in ESVS is taken in the further processing stage: fusion of radar and IR data hypotheses and synthetic vision data. The functionality of the algorithm was tested extensively during several flight campaigns with landing approaches to different German and European airports in 2003 and 2004.

The question of whether there is a preferred or best classifier to use with remotely sensed data is discussed, focussing on likely results and ease of training. By appealing in part to the No Free Lunch Theorem, it is suggested that there is really no superiority of one well trained algorithm over another, but rather it is the means by which the algorithm is employed - ie. the classification methodology - that often governs the outcomes.

Technological advances in digital imaging and liquid crystal tunable filters allow for a design of a both compact and cost effective multispectral camera system. Thus an image acquisition system for the visible and near infrared wavelength range can be realised consisting of conventional CCD cameras and tunable filters whose spectral transmittace is controlled electronically. Due to their limited internal payload the use of a lightweight, compact camera system is of particular importance for applications involving mini-unmanned-aerial-vehicles (mini-UAVs). Within the scope of this paper a both compact and economical data acquisition system for multispecral images is described. Despite of their limited functionality (e.g.regarding calibration) in comparison with commercial systems such as AVIRIS the use of these upcoming compact multispectral camera systems can be advantageous in many applications. In order to maintain the systems low weight and price this work proposes to separate data acquisition and processing modules, and transmit pre-processed camera data online to a stationary high performance computer for further processing. Special focus is laid upon transmission paths with limited bandwidth.

In global classifications using Markov Random Field (MRF) modelling, the neighbourhood form is generally considered as independent of its location in the image. Such an approach may lead to classification errors for pixels located at the segment borders. The solution proposed here consists in relaxing the assumption of fixed-form neighbourhood. However this non-stationary neighbourhood modelling is useful only if an efficient heuristic can be defined to perform the optimization. Ant colony optimization (ACO) is currently a popular algorithm. It models upon the behavior of social insects for computing strategies: the information gathered by simple autonomous mobile agents, called ants, is shared and exploited for problem solving. Here we propose to use the ACO and to exploit its ability of self-organization. The ants collect information through the image, from one pixel to the others. The choice of the path is a function of the pixel label, favouring paths within a same image segment. We show that this corresponds to an automatic adaptation of the neighbourhood to the segment form. Performance of this new approach is illustrated on a simulated image and on actual remote sensing images, SPOT4/HRV, representing agricultural areas. In the studied examples, we found that it outperforms the fixed-form neighbourhood used in classical MRF classifications. The advantage of having a neighborhood shape that automatically adapts to the image segment clearly appears in these cases of images containing fine elements, lanes or thin fields, but also complex natural landscape structures.

Spatial AdaBoost, proposed by Nishii and Eguchi (2005), is a machine learning technique for contextual supervised image classification of land-cover categories of geostatistical data. The method classifies a pixel through a convex combination of a log posterior probability at the current pixel and averages of log posteriors in various neighborhoods of the pixel. Weights for the log posteriors are tuned by minimizing the empirical risk based on the exponential loss function. It is known that the method classifies test data very fast and shows a similar performance to the Markov-random-field-based classifier in many cases. However, it is also known that the classifier gives a poor result for some data when the exponential loss puts too big penalty for misclassified data. In this paper, we consider a robust Spatial boosting method by taking a robust loss function instead of the exponential loss. For example, the logit loss function gives a linear penalty for misclassified data approximately, and is robust. The Spatial boosting methods are applied to artificial multispectral images and benchmark data sets. It is shown that Spatial LogitBoost based on the logit loss can classify the benchmark data very well even though Spatial AdaBoost based on the exponential loss failed to classify the data.

This paper describes a new method for classification of hyperspectral images for extracting carthographic objects. The developed method is intended as a tool for automatic map updating. The idea is to use an existing map of the region of interest as a learning set. The proposed method is based on logistic regression. Logistic regression (LR) is a supervised multi-variate statistical tool that finds an optimal combination of the input channels for distinguishing one class from all the others. LR thus results in detection images per class. These can be combined into a classification image. The LR method that is used here is a step-wise optimisation that also performs a channel selection. The results of the LR are further improved by taking into account spatial information by means of a region growing method. The parameters of the region growing are optimised for each class of interest. For each class the optimal set of parameters is determined. The method is applied on a HyMap hyperspectral image of an area in Southern Germany and the results are compared to those of classical methods. For the comparison a ground truth image was created by combining data from a cadaster map and a digital topographic map.

Hyperspectral image classification impose challenging requirements to a classifier. It is well known that more spectral bands can be difficult to process and introduce problems such as the Hughes phenomenon. Nevertheless, user requirements are very demanding, as expectations grow with the available number of spectral bands: subtle differences in a large number of classes must be distinguished. As multiclass classifiers become rather complex for a large number of classes, a combination of binary classification results are often used to come to a class decision. In this approach, the posterior probability is retained for each of the binary classifiers. From these, a combined posterior probability for the multiclass case is obtained. The proposed technique is applied to map the highly diverse Belgian coastline. In total, 17 vegetation types are defined. Additionally, bare soil, shadow, water and urban area are also classified. The posterior probabilities are used for unmixing. This is demonstrated for 4 classes: bare soil and 3 vegetation classes. Results are very promosing, outperforming other approaches such as linear unmixing.

This paper introduces a semisupervised classification method, which exploits both labeled and unlabeled samples, for addressing "ill-posed" problems with support vector machines (SVMs). The method is based on recent developments in statistical learning theory concerning transductive inference and in particular Transductive SVMs (TSVMs). We propose a novel modified TSVM classifier designed for the analysis of "ill-posed" remotesensing problems. In particular, the proposed technique: i) is based on a novel transductive procedure that exploits a weighting strategy for the unlabeled patterns based on a time-dependent criterion; ii) is developed also for multiclass cases; and iii) addresses the model-selection problem with lack of test/validation sets. Experimental results confirm the effectiveness of the proposed method on a set of "ill-posed" remote-sensing classification problems representing different operative conditions.

This paper presents a method for combining both spectral and spatial features to perform hyperspectral image classification. Texture based spatial features computed from statistical, wavelet multiresolution, Fourier spectrum and Gabor filters are considered. A step wise feature selection method selects optimal set of features from the combined feature set. A comparison of the different spatial features for improving hyperspectral image classification is presented. The results show that wavelet based features and statistical features perform best. The effect of band subset selection using information based subset selection methods on the combined feature set is presented. Several results with hyperspectral images show the efficacy of utilizing spatial features.

ICA has been a well studied subject in recent years. Its implementation may employ neural networks or adaptive learning techniques. In contrast to PCA, the objective of ICA is to extract components with high-order statistical independence. The concept and process of deriving the independent components have had motivated the development of many mathematical algorithms. In fact it is not necessary to achieve perfect statistical independence in this process. ICA is particularly, perhaps uniquely also, useful in blind source separation problem which is to determine from the received signals the original signals from different physical sources which are considered as independent. ICA has significant impact on many applications such as in remote sensing, medical testing, face recognition, direction of arrival estimatin and other areas The purpose of this paper is to examine some of these applications including SAR images, sonar signals, and exploration seismic data.

Dimensionality reduction plays a crucial role in many hyperspectral data processing and analysis algorithms. This paper proposes a new mean squared error based approach to determine the signal subspace in hyperspectral imagery. The method first estimates the signal and noise correlations matrices, then it selects the subset of eigenvalues that best represents the signal subspace in the least square sense. The effectiveness of the proposed method is illustrated using simulated and real hyperspectral images.

Using several hyperspectral images acquired by various aerospace sensors over San Rossore (Italy) test site, different issues concerning the assessment of the noise amplitude, which affects the remotely sensed images, are investigated. An innovative algorithm, developed by the authors, is presented and its performance is discussed. This procedure analyses the bit-planes extracted from any monochromatic image in the hyperspectral data cube, then it assesses the randomness of every bit-plane, and computes the signal-to-noise ratio for each spectral channel. Differently from more traditional signal-to-noise ratio estimators, which need an homogeneous area in the concerned image to isolate noise contribution only, the new algorithm is almost insensitive to scene texture. Due to this property the developed method is able to process the image of any observed ground area. The paper discusses other possible sources of systematic disturbance (like stripe-noise, smear effect, and so forth), which may dim the quality of remotely sensed data. Finally, the behaviour of signal-to-noise ratio of CHRIS and other hyperspectral sensors like MIVIS and VIRS-200 is shown as function of wavelengths.

Anomaly detectors reveal the presence of objects/materials in a multi/hyperspectral image simply searching for those pixels whose spectrum differs from the background one (anomalies). This procedure can be applied directly to the radiance at the sensor level and has the great advantage of avoiding the difficult step of atmospheric correction. The most popular anomaly detector is the RX algorithm derived by Yu and Reed. It is based on the assumption that the pixels, in a region around the one under test, follow a single multivariate Gaussian distribution. Unfortunately, such a hypothesis is generally not met in actual scenarios and a large number of false alarms is usually experienced when the RX algorithm is applied in practice. In this paper, a more general approach to anomaly detection is considered based on the assumption that the background contains different terrain types (clusters) each of them Gaussian distributed. In this approach the parameters of each cluster are estimated and used in the detection process. Two detectors are considered: the SEM-RX and the K-means RX. Both the algorithms follow two steps: first, 1) the parameters of the background clusters are estimated, then, 2) a detection rule based on the RX test is applied. The SEM-RX stems from the GMM and employs the SEM algorithm to estimate the clusters' parameters; instead, the K-means RX resorts to the well known K-means algorithm to obtain the background clusters. An automatic procedure is defined, for both the detectors, to select the number of clusters and a novel criterion is proposed to set the test threshold. The performances of the two detectors are also evaluated on an experimental data set and compared to the ones of the RX algorithm. The comparative analysis is carried out in terms of experimental Receiver Operating Characteristics.

Harmful algal blooms are believed to be increasing in occurrence and their toxins can be concentrated by filter-feeding shellfish and cause amnesia or paralysis when ingested. As a result fisheries and beaches in the vicinity of blooms may need to be closed and the local population informed. For this avoidance planning timely information on the existence of a bloom, its species and an accurate map of its extent would be prudent. Current research to detect these blooms from space has mainly concentrated on spectral approaches towards determining species. We present a novel statistics-based background-subtraction technique that produces improved descriptions of an anomaly's extent from remotely-sensed ocean colour data. This is achieved by extracting bulk information from a background model; this is complemented by a computer vision ramp filtering technique to specifically detect the perimeter of the anomaly. The complete extraction technique uses temporal-variance estimates which control the subtraction of the scene of interest from the time-weighted background estimate, producing confidence maps of anomaly extent. Through the variance estimates the method learns the associated noise present in the data sequence, providing robustness, and allowing generic application. Further, the use of the median for the background model reduces the effects of anomalies that appear within the time sequence used to generate it, allowing seasonal variations in the background levels to be closely followed. To illustrate the detection algorithm's application, it has been applied to two spectrally different oceanic regions.

Change detection is an important part of many remote sensing applications. This paper addresses the problem of unsupervised pixel classification into 'Change' and 'No Change' classes based on Hidden Markov Random Field (HMRF) models. HMRF models have long been recognized as a method to enforce spatially coherent class assignment. The optimal classification under these models is usually obtained under the Maximum a Posteriori (MAP) criterion. However, the MAP classification in HMRF models leads in general to problems with exponential complexity, so approximate techniques are needed. In this paper we show that the simple structure of the change detection problem makes that MAP classification can be exactly and efficiently calculated using graph cut techniques. Another problem related to HMRF modelling (and to any change detection technique) is the determination of the parameters or thresholds for classification. This learning problem is solved in our HMRF model by the Expectation Maximization (EM) algorithm. Experimental results obtained on four sets of multispectral remote sensing images confirm the validity of the proposed approach.

The detection of changes affecting continental surfaces has important applications in hydrological, meteorological, and climatic modelling, so that numerous change indices have already been proposed that use remote sensing data. In this work, we show the interest of combining several of them to improve change detection performance. The combination is done in the Dempster-Shafer evidence theory framework, therefore allowing ignorance modelling. Each mass function is defined either based on the result of the corresponding mono-index analysis, that is done using an 'a contrario' approach, or from generic sigmoid function in the absence of pdf assumption. Using actual SPOT/HRV data, we analyse the performance of different change indices, and their combination in different application cases: forest fires, forest logging either in pine forest or in mixed forest, and winter vegetation cover of fields in intensive farming areas. Finally, we also show the interest of the indices derived from the Information Theory, some of which being original.

At the German Aerospace Center, DLR, an automatic and operational traffic processor for the TerraSAR-X ground segment is currently under development. The processor comprises the detection of moving objects on ground, their correct assignment to the road network, and the estimation of their velocities. Since traffic flow parameters are required for describing dynamics and efficiency of transportation, the estimation of the velocity of detected ground moving vehicles is an important task in traffic research science. In this paper we show for TerraSAR-X simulated data how the along-track velocity component of a moving vehicle can be derived indirectly by processing SAR data with varying frequency modulation (FM) rates and exploiting the specific behavior of the vehicle's signal through the FM rate space. An airborne ATI-SAR campaign with DLR's ESAR sensor has been conducted in April 2004 in order to investigate the different effects of ground moving objects on SAR data and to acquire a data basis for algorithm development and validation. Several test cars equipped with GPS sensors as well as vehicles of opportunity on motorways with unknown velocities were imaged with the radar under different conditions. To acquire reference data of superior quality, all vehicles were simultaneously imaged by an optical sensor on the same aircraft allowing their velocity estimation from sequences of images. The paper concentrates on the estimation of along-track velocities of moving vehicles from SAR data. Velocity measurements of vehicles in controlled experiments are presented, including data processing, comparison with GPS and optical reference data and error analysis.

In this paper, we propose an extension of the automatic and unsupervised change-detection method presented in [1]. By analyzing a properly defined cost function, the proposed method allows the automatic detection of the number (zero, one or two) and the values of the decision thresholds associated with changes (if any) in the log-ratio image. This cost function is the minimum value of the criterion function adopted to select the decision threshold in the log-ratio image according to a modified double thresholding Kittler Illingworth (KI) algorithm (implemented under the Generalized Gaussian (GG) assumption for changed and unchanged classes). Experimental results carried out both on real and simulated multitemporal SAR images proved the effectiveness of the proposed automatic method in detecting both the number of changes (the situation of no changes is also explicitly identified) and the related threshold values.

A comparison of different classification approaches for multitemporal SAR images data sets is provided in this work. The aim is to assess the performance of estimators of the backscatter temporal variability in terms of classification accuracy for a typical four-class problem. Different approaches in forming an appropriate feature vector are discussed and compared with multichannel classifiers like the fuzzy k-means. Finally, a classifier that employs a feature fusion step based on principal components analysis is proven promising since it provides increased classification accuracy and reduced computational complexity.

A novel iterative approach based on a modular neural architecture (Jacobs, Jordan, and Barto, 1990) is presented for the classification of SAR images of sea ice. In addition to the local image information, the algorithm uses spatial context information derived from the first iteration of the algorithm and refines it in the subsequent iterations. The modular structure of the neural network aims to capture structural features of the SAR images of sea ice in the Marginal Ice Zone.

This work focuses on estimating the information conveyed to a user by multi-dimensional digitised signals. The goal is establishing the extent to which an increase in radiometric resolution, or equivalently in signal-to-noise-ratio (SNR), can increase the amount of information available to users. Lossless data compression and noise modeling are exploited to measure the "useful" information content of the data. In fact, the bit-rate achieved by the reversible compression process takes into account both the contribution of the "observation" noise, i.e. information regarded as statistical uncertainty, whose relevance is null to a user, and the intrinsic information of hypothetically noise-free samples. Once the parametric model of the noise, assumed to be possibly non-Gaussian, has been preliminarily estimated, the mutual information between noise-free signal and recorded noisy signal is easily estimated. However, it is also desirable to know what is the amount of information that the digitised samples would convey if they were ideally recorded without observation noise. Therefore, an entropy model of the source is defined and such a model is inverted to yield an estimate of the information content of the noise-free source from the code rate and the noise model. Results are reported and discussed on true superspectral data (14 spectral bands) recorded by the ASTER imaging radiometer.

Multichannel (multispectral) remote sensing (MRS) is widely used for various applications nowadays. However, original images are commonly corrupted by noise and other distortions. This prevents reliable retrieval of useful information from remote sensing data. Because of this, image pre-filtering and/or reconstruction are typical stages of multichannel image processing. And majority of modern efficient methods for image pre-processing requires availability of a priori information concerning noise type and its statistical characteristics. Thus, there is a great need in automatic blind methods for determination of noise type and its characteristics. However, almost all such methods fail to perform appropriately well if an image under consideration contains a large percentage of texture regions, details and edges. In this paper we demonstrate that by applying bootstrap it is possible to obtain rather accurate estimates of noise variance that can be used either as the final or preliminary ones. Different quantiles (order statistics) are used as initial estimates of mode location for distribution of noise variance local estimations and then bootstrap is applied for their joint analysis. To further improve accuracy of noise variance estimations, it is proposed under certain condition to apply myriad operation with tunable parameter k set in accordance with preliminary estimate obtained by bootstrap. Numerical simulation results confirm applicability of the proposed approach and produce data allowing to evaluate method accuracy.

In this communication, we evaluate the performance of the relevance vector machine (RVM) (Tipping,2000) for the estimation of biophysical parameters from remote sensing images. For illustration purposes, we focus on the estimation of chlorophyll concentrations from multispectral imagery, whose measurements are subject to high levels of uncertainty, both regarding the difficulties in ground-truth data acquisition, and when comparing in situ measurements against satellite-derived data. Moreover, acquired data are commonly affected by noise in the acquisition phase, and time mismatch between the acquired image and the recorded measurements, which is critical for instance for coastal water monitoring. In this context, robust and stable regressors that provide inverse models are desirable. Lately, the use of the support vector regressor (SVR) has produced good results to this end. However, the SVR has many deficiencies, which could be theoretically alleviated by the RVM. In this paper, performance of the RVM is benchmarked with SVR in terms of accuracy and bias of the estimations, sparseness of the solutions, distribution of the residuals, robustness to low number of training samples, and computational burden. In addition, some theoretical issues are discussed, such as the sensitivity to hyperparameters setting, kernel selection, and confidence intervals on the predictions. Results suggest that RVM offer an excellent compromise between accuracy and sparsity of the solution, and reveal itself as less sensitive to selection of the free parameters. Some disadvantages are also pointed, such as the unintuitive confidence intervals provided and the computational cost.

In the present paper is studied the effect of a recently proposed filter on satellite images. In spatial frequency domain, the response of the filter is controlled by the parameters b and k, which are positive real numbers. 5 x 5 convolution masks at the image are constructed, which simulate the response of the filter at frequency domain, for various b and k values. In order to study the performance of the filter in qualitative and quantitative terms, these masks were applied on various satellite images, which were taken over urban centers, as well as broader regions with different land cover types and geomorphological features. For each filtered image, there were computed the standard deviation and the signal to noise ratio. The statistical analysis showed that for k between 0.5 and 0.7 and for b between 0.8 and 1.2, one can produce images with considerably enhanced tonalities between adjacent pixels. Experimentation with satellite images showed that convolution masks with these k and b parameters produce images where lineaments such as coastal lines, drainage or road network can be clearly seen. The potential user is encouraged to try on various k and b values, in order to obtain the optimum result for the area under study.

Spectral unmixing decomposes an hyperspectral image into a collection of reflectance spectra of the macroscopic materials present in the scene, called endmembers, and the corresponding abundance fractions of these constituents. The purpose of this paper is to compare the performance of several algorithms that process unsupervised endmember extraction from hyperspectral images in the visible and NIR spectral ranges. After giving an analytical formulation of the observations, two significantly different approaches have been described. The first one exploits convex geometry the problem answers to. The second one is based on statistical principles of Independent Component Analysis, which is a classical resolution of the Blind Source Separation issue. First, the performance of the algorithms are compared on synthetic images and sensibility to noise is studied. Then the best methods are applied on part of a HyMap image.

Location and characterisation of magnetic objects from measured magnetic data has long been a research interest with the difficulty in handling the non-uniqueness in the inversion process. Ground-surface methods, which are widely used for objects buried in shallow depthes, become ineffective for those sinking into deep soil, because the anomaly field diminishes rapidly by distance and is heavily interfered by the metal debris distributed in the ground surface. A total-field borehole magnetometer can penetrate to such depths and collect relatively quiet data. However, conventional interpretation techniques suffer from the inherent non-uniqueness in the borehole dimension. In this paper, a constrained optimisation method is utilised for the interpretation of total-field borehole data for the detection of deeply buried objects. The major advantage over conventional techniques comes from the analytically-derived constraints imposed on the parameter vector by excluding the non-geosensible results from consideration and hence reducing the non-uniqueness to a minimal level. A test site has been developed for the evaluation using real world data. The interpretation results demonstrate its superior capability in handling real-world problems with high non-uniqueness. Furthermore, this method provides a way to estimate the moment strength without knowing the exact position. Together with the modelled signature data for different objects, characterisation of a particular item is possible from the inversion of a single total-field borehole profile.

This paper addresses the problem of sub-pixel target detection in hyperspectral images assuming that the target spectral signature is deterministic and known. Hyperspectral image pixels are frequently a combination or mixture of disparate materials or components. The need of a quantitative pixel decomposition arises in many civilian and military applications such as material classification, anomaly and target detection. The Linear Mixing Model (LMM) is a widely used method in hyperspectral data analysis. It represents a mixed pixel as the sum of the spectra of known pure materials, called endmembers, weighted by their relative concentrations called abundance coefficients. However, the LMM does not take into account the natural spectral variability of the endmembers. This variability is well represented by the Stochastic Mixing Model (SMM), which provides a model for describing both mixed pixels in the scene and endmember spectral variations through a statistical model. Modeling the background spectrum as a Gaussian random vector with known mean spectrum and unknown covariance matrix, a novel SMM based Detector (ASMMD) is derived in this paper. The ASMMD theoretical performances are evaluated in a case study referring to actual conditions. The analysis is conducted by estimating the ASMMD parameters on an experimental data set acquired by the AVIRIS hyperspectral sensor and the results are compared with the ones achieved by the Adaptive Matched Subspace Detector (AMSD), based on the LMM.

A new sensor-specific correlated-k (c-k) ultra-fast radiative transfer (RT) formalism, kURT, has been designed for fast broad-bandpass scene simulations from UV-visible to LWIR wavelengths. A higher resolution RT code (1 cm^-1 MODTRAN) has been adapted to output 1 cm^-1 correlated-k parameters for ozone, water, and the combined uniformly mixed species on a pressure-temperature grid, which are merged to form a compact c-k set incorporating the sensor bandpass response function. The compact set is used to compute bandpass transmittance and radiance in near-real time. Scattering parameters (molecular Rayleigh, clouds and aerosols), blackbody and solar functions are cast as compact k-dependent source terms and used in the radiance computations. Preliminary transmittance results for 3-5 and 8-12 micron bandpasses and visible-MWIR sensors yield results within 2% of a 1 cm^-1 MODTRAN calculation with a two-orders-of-magnitude computational savings. Applications include near-earth broadband propagation and extinction calculations for target detection and recognition, mid-range tracking, and search and rescue operations from ground and low altitude aircraft.

At the moment of carry out a study with ground penetrating radar (GPR) it is interesting to count with the support provided by other information sources. All the available information relative to the study area will be valuable in the subsequent phases of processing and interpretation of the obtained GPR records. Nowadays there is a logical trend to the integration of GPS devices. The decrease in size of these equipment, the increase of their accuracy and new wireless communication technologies (802.11, Bluetooth,...) encourage this incorporation. GPR/GPS integration allows an accuracy positioning of the radar data under favourable conditions. Furthermore it brings the possibility to import this data into a geographic information system (GIS). This study deepens the process of integration of both technologies applied to road evaluation. To the accomplishment of this study, a dual frequency (L1+L2) RTK GPS, two Bluetooth GPS receivers (with SiRF chip) admitting both real time differential corrections (SBAS), and a GPS receiver with post-processed sub-meter accuracy were used. As regards GPR equipment, shielded 500, 800 and 1000 MHz antennas were used in different configurations.

This paper deals with the detection problem of non-metallic anti-personnel (AP) land mines by using ground penetrating radar (GPR). The clutter signal due to the reflection from the ground surface is often a performance limiting factor in GPR. An algorithm has been developed for this purpose, which combines two powerful tools: The wavelet packet analysis and the higher-order-statistics (HOS). The use of both techniques makes the detection of shallow AP land mines objects, which are obscured by the return from air-soil interface, possible. A de-noising procedure is performed on the wavelet coefficients and the classical method of detection is then applied to these coefficients. The used threshold is based on the higher order statistics. The algorithm has been tested using two source data, one-GHz pulse GPR data, and one-to-four-GHz stepped-frequency GPR data, from laboratory measurements.

In this paper, a new approach for unsupervised change-detection using multitemporal InSAR data is proposed, of which the significant characteristics is joint use of backscattering temporal intensity and long-term coherence based on 2-D (two dimensional) Renyi's entropy. The proposed approach is made up of two steps: feature extraction and unsupervised 2-D thresholding. In the first step, two features are based on the concepts of backscattering intensity variation and long-term coherence variation respectively, and are defined according to the analysis of different signal behavior of interferometric SAR in the presence of land-cover classes within urban area. In the second step, an unsupervised 2-D thresholding technique based on maximum 2-D Renyi's entropy criterion is developed. The thresholding is performed on the two difference images derived from the two features to produce an accurate change-detection map with two classes: changed and no-changed. Primary experimental results, which were obtained from a set of six multitemporal ERS-1/2 SAR images within Shanghai city of China, show the effective of the proposed approach and that ERS-1/2 InSAR data could be exploited for detecting urban land-cover changes.

A simple and fast multi-channel filtering algorithm is presented in the paper for texture segmentation. This algorithm requires only a small number of channels and automatically determines the channel parameters by analysis of the Fourier power spectrum. Also it does not need a priori knowledge about the type and number of textures occurring in the input images. We can gain categories number by the square-error plot of different clusters. It does not involve any human intervention. The algorithm is tested extensively with a variety of Brodatz's textures and real textures. The segmentation results validate the practicability of this algorithm.

Image fusion aims at the exploitation of the information conveyed by data acquired by different imaging sensors. A notable application is merging images acquired from space by panchromatic and multi- or hyper-spectral sensors that exhibit complementary spatial and spectral resolution. Multiresolution analysis has been recognized efficient for image fusion. The Generalized Laplacian Pyramid (GLP), in particular, has been proven as the most efficient scheme due to its capability of managing images whose scale ratios are fractional numbers (non-dyadic data) and to its simple and easy implementation. Data merge based on multiresolution analysis, however, requires the definition of a model establishing how the missing spatial details to be injected into the multi-spectral bands are extracted from the panchromatic image. The model can be global over the whole image or depend on the local space-spectral context. This paper reports results on the fusion of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) data. Each of the five thermal infrared (TIR) images (90m) is merged with the most correlated visible-near infrared (VNIR) image (15m). Due to the 6:1 scale ratio, the GLP has been utilized. The injection of spatial details has been ruled by means of the Spectral Distortion Minimizing (SDM) model that minimizes the spectral distortion between the resampled and fused images. Notwithstanding the lack of a spectral overlap between the VNIR and the TIR bands, experimental results show that the fused images keep their spectral characteristics while the spatial resolution is enhanced.