High resolution imaging is a persistent goal for many users of optical and microwave remote sensing data as it allows better detection, discrimination and classification of the objects contained within a recorded scene. High resolution can be obtained either by instrument design, by data acquisition methods, or by dedicated post-processing and interpretation of the acquired data. The interpretation can be based on a visual inspection by a human photointerpreter, or consist of automated model-based scene understanding. We will compare various methods used for optical and SAR instruments and their processing chains. Primary criteria for high resolution imaging as encountered in remote sensing of solid surfaces are ground resolution per pixel, motion compensation during data acquisition, attainable contrast and signal-to-noise ratio, removal of instrumental and non-target effects, geometrical correction, use of multi-channel and neighborhood target property data like spectral and textural signatures, and the potential for data fusion from multiple sources. The inclusion of model knowledge obtained from collections of pre-recorded physical target data leads to a comparison of the acquired data with representative models. Similarities and deviations revealed during the comparison allow a detailed high resolution interpretation of the image data and lead to a full image understanding.

This paper demonstrates improvements in GOES on-orbit Image Navigation and Registration (INR) performance obtained from a careful method of landmark selection, coupled with changes in landmark measurement techniques. An iterative process was used which began with a search for characteristics influencing landmark stability over time, such as topography, thermal changes, and apparent sea/land boundaries, and ended with measurable improvements in the areas of navigation, within- frame, frame-to-frame, and channel-to-channel registration. An operational shift from manual to automatic landmark correlation proved to be an essential requirement influencing the overall results. Specification requirements are used as an absolute means to estimate INR performance changes for GOES 8, but whenever possible, results obtained here are also placed in perspective by comparison with previous GOES INR results. Finally, the methodology and tools described are applied to the new GOES 10 satellite in order to estimate its performance.

The purpose of this study is to improve mathematical modeling of calibration curves produced by the Landsat calibrators. We explain one band and one lamp modeling and then one calibration band and multiple lamps averaging. The algorithm has three parts at the present, namely one dimensional modeling that includes a change-point removal and two or more signals averaging. A demo of the algorithm and the data is available from the Internet using any Web browser.

There are two major approaches in spectral unmixing: linear and non-linear ones. They are appropriate for different types of mixture, namely checkerboard mixtures and intimate mixtures. The two approaches are briefly reviewed. Then in a carefully controlled laboratory experiment, the limitations and applicability of two of the methods (a linear and a non- linear one) are compared, in the context of unmixing an intimate mixture.

We present a method for the automatic, unsupervised detection of spectrally distinct targets from the background using hyperspectral imaging. The approach is based on the concepts of projection pursuit (PP) and unsupervised orthogonal subspace projection (UOSP). It has the advantage of not requiring any prior knowledge of the scene or the objects' spectral signatures. All information is obtained from the data. First, PP is used to both reduce the data dimensionality and locate potential targets. Then, UOSP suppresses the signatures from undesired objects or interferers that cause false detections when a spectral filter is applied. The result is a set of gray scale images where objects belonging to the same spectral class are enhanced while the background and other undesired objects are suppressed. This method is demonstrated using data from the Hyperspectral Digital Imagery Collection Experiment (HYDICE).

The problem of unsupervised clustering of data is formulated using a Bayesian inference. The entropy is considered to define a prior. In clustering problem we have to reduce the complexity of the gray level description. Therefore we minimize the entropy associated with the clustering histogram. This enables us to overcome the problem of defining a priori the number of clusters and an initialization of their centers. Under the assumption of a normal distribution of data the proposed clustering method reduces to a deterministic algorithm (very fast) which appears to be an extension of the standard k-means clustering algorithm. Our model depends on a parameter weighting the prior term and the goodness of fit term. This hyper-parameter allows us to define the coarseness of the clustering and is data independent. Heuristic argument is proposed to estimate this parameter. The new clustering approach was successfully tested on a database of 65 magnetic resonance images and remote sensing images.

A new sub-optimal search strategy suitable for feature selection in high-dimensional remote-sensing images (e.g. images acquired by hyperspectral sensors) is proposed. Such a strategy is based on a search for constrained local extremes in a discrete binary space. In particular, two different algorithms are presented that achieve a different trade-off between effectiveness of selected features and computational cost. The proposed algorithms are compared with the classical sequential forward selection (SFS) and sequential forward floating selection (SFFS) sub-optimal techniques: the first one is a simple but widely used technique; the second one is considered to be very effective for high-dimensional problems. Hyperspectral remote-sensing images acquired by the AVIRIS sensor are used for such comparisons. Experimental results point out the effectiveness of the presented algorithms.

Orthogonal subspace projection (OSP) has been successfully applied to hyperspectral image processing. In order for OSP to be effective, the number of bands must be no less than that of signatures to be classified so that there are sufficient dimensions to accommodate individual signatures to discriminate one another via orthogonal projection. This intrinsic constraint is not an issue for hyperspectral images since they generally have hundreds of bands which are more than the number of signatures resident within images. It, however, may not be true for multispectral images where the number of signatures to be classified is greater than the number of bands such as 3-band SPOT images. This paper presents a generalization of OSP, called generalized OSP (GOSP) to relax this constraint in such a fashion that OSP can be extended to multispectral image processing in an unsupervised fashion. The idea of GOSP is to create new additional band images nonlinearly from original multispectral images so as to achieve sufficient dimensionality prior to OSP classification. It is then followed by an unsupervised OSP classifier, called automatic target detection and classification algorithm (ATDCA) for classification. The effectiveness of the proposed GOSP is evaluated by a 3-band SPOT and a 4-band Landsat MSS images. The experimental results has shown that GOSP significantly improves the classification performance of OSP.

The speckle size metric known as power-averaged speckle size (PASS), which is based on the integral of the power spectral density of the data, is insensitive to target asymmetry. PASS is defined as the inverse of the median frequency of the power spectral density of the reconstructed pupil. Implementation of the metric using simulated data from different targets under a variety of imaging conditions illustrates the impact of the environment. In this paper, we deconvolve the imaging environment using power spectral density of the detected speckle in the error function. Different targets are simulated in the presence of a variety of imaging situations as well as noise and we compare the resultant deconvolved images with undistorted images as well as with the unimproved ones.

In Alenia Aerospazio has been developed a demonstrator for SAR data analysis and automatic target recognition (ATR). The new version we present in this paper introduces some new characteristics. A better sidelobes suppression via the locally variant apodization and a new CFAR processor using a cell-average or an ordered-statistics detector are integrated to the pervious structure to improve the preliminary target detection i.e. the candidate ship detection. After this step and in parallel with the ship wake detection we implement a new and independent way to recognize moving objects with a following improvement in detection performances: the estimation of residual coherent phase error locally present on the focused image, due to the target motion. The test images are referred to the SIR-C/X-SAR mission.

A wavelet-based algorithm for polarimetric SAR imagery segmentation is developed. It utilizes the property that under wavelet transform correlation is very weak among the intensity HH, HV, VV channels as well as among the subimages in the same channel to form the effective feature vector for parametric segmentation. The statistical separability measures including the Bhattacharyya distance and a separability inhomogeneity function (SIF) are employed to extract the most effective feature vector in the sense of minimizing SIF. The algorithm is applied to the supervised segmentation of the polarimetric SAR imagery of San Francisco Bay area. It shows a good segmentation performance and a significant computational reduction. The segmentation result is also favorably compared with that of Gaussian modeling or mixture Gaussian modeling under non-feature extraction representation.

We address the problem of edge detection in synthetic aperture radar (SAR) images and we particularly focus on the precision with which the frontier of an object can be determined. Recently, an edge detection global filter has been proposed. It involves a two-region analyzing window whose properties greatly influence its efficiency. In this paper, we will study the filter detection and localization abilities as functions of the analyzing window geometry, namely its shape, size and orientation. We will then introduce another approach to segment an object in a SAR image: the statistical active contour. The spatial precision of these two approaches will be finally compared and discussed.

Discrimination of vegetation canopies for production of forestry and land use thematic cartography from multispectral satellite images requires high spectral and spatial resolutions, usually not available in this type of images. A methodology is proposed to improve a vegetation oriented classification from a Landsat TM image by adding texture information obtained from panchromatic aerial photographs. Multispectral classification was used to create a mask of the forested areas that was applied over the aerial mosaic composition. Further vegetation classes were defined based on textural differences, and eight texture features derived from the gray level co-occurrence matrix, three textural energy indicators and a factor of edgeness were tested. A selection of optimal features and textural parameters such as number of gray levels, window size and distance between pixels was performed using principal components and stepwise discriminant analysis techniques with a set of representative samples from each class. After a texture segmentation of panchromatic aerial imagery using optimal parameters and features was completed, a post-classification process based on morphological operations was applied to avoid the neighboring effect generated by the texture analysis. Overall accuracy in the identification of texture classes using the four best feathers was 86.6%, while the 88% of accuracy was achieved in the classification of the complete image. This method is useful for discrimination of certain vegetation classes with low spectral separability and arranged in small forest units, increasing the classification detail in those areas of particular interest.

Local-statistics speckle filtering has been extended to multitemporal SAR data by exploiting the temporal correlation of the speckle noise across a set of images of the same scene taken at different times. A recursive nonlinear transformation aimed at decorrelating the data across time, while retaining the multiplicative noise model, is defined from the geometric means and the ratios of couples of spatially overlapped observations. The temporal correlation coefficient (TCC) is estimated from the modes of the distributions of the local variation coefficient C_v computed on transformed couples of images. The images are filtered in the transformed domain and reversely transformed to yield despeckled observations in which seasonal changes are preserved, or even highlighted, and texture analysis is expedited. Tests on four SAR images from repeat-pass ERS-1 corroborate the theoretical assumptions and show the filtering performances of the proposed approach.

SAR images from the ERS satellites have proved to be helpful data for identification of oil spills. Because the presence of oil slicks on sea surface increases the surface tension of sea water, the surface wave motion is significantly depressed. This effect relatively reduces the sea surface roughness, decreases the radar backscattered energy and enables oil slicks to be discernible from the radar image. The use of fractal dimension, which is related to the concept of surface 'roughness,' as a feature for classification, improves the oil spill detection, since enhances texture discrimination with respect to first and second order derivative operators, e.g., DoG and LoG. This paper describes a multi-resolution approach based on fractal geometry for oil spill detection in ERS SAR images. The proposed multi-resolution algorithm is based on the normalized Laplacian pyramid which provides a band-pass description of the image. Thanks to normalization of each layer of the pyramid by its low-pass version, the image noise becomes independent on the image signal, and a reliable estimate of the fractal dimension can be computed from ratio of power spectra at different scales. The experimental results carried out both on synthetic and ERS-1 SAR images prove the effectiveness of the fractal-based approach for the classification of oil spills.

The goal of image segmentation is to process the data given by the pixels so as to obtain a 'meaningful' partitioning of the image. The pixels of an image are represented by points in an n-dimensional space (spectral space). There is no universally accepted definition for texture, here texture means relation among pixels, given by the statistic of the image. In this work, an algorithm has been developed for segmenting color images (or multispectral), after performing a clustering over the set of points in spectral space. The algorithm captures texture information in order to segment the image. Finally, results obtained with synthetic and real aerial images are shown.

Screening and thematic data exploitation are important tasks in the reconnaissance cycle. To reduce the work load of the image analysts, an automatization of these tasks is required. For an automatic image exploitation structural image analysis algorithms were developed for multisensor data. The approach is based on the primitive objects line which are generated by appropriate edge detectors. For preprocessing in SAR images, i.e. the change from the iconic to the symbolic level, not only speckle noise prejudices automatic segmentation but also multiple scattering of hard targets in the scene. Several different approaches to filter speckle noise and to detect edges prior to structural image analysis have been investigated. The structural analysis of complex scenes uses the blackboard-based production system (BPI) as framework. The primitive objects are built up step by step applying the productions. For an IMINT report the image data has to be referenced to a cartographic grid. In an automated reconnaissance cycle an automatic image-to-map registration is required. The necessary control points must be detected in the image data and a correspondence between map and image control points must be found. The transformation parameters can then be calculated. Additionally, when using map information, expectation areas can be defined and processing needs can be reduced efficiently.

In order to improve the correct classification rate of the conventional maximum likelihood method for classification of multi-spectral images, we introduce 'a priori probability' estimated from the spatial structure of the images. In this we consider the observed data as random field defined on a 2D lattice. Each pixel has a class label which is also regarded as random field on the lattice. Then the spatial structure of an image is expressed by the dependence of a label on its neighbors. We use local and global spatial information of an image in classification process by making a point in the label lattice have both local and global interrelations. To accomplish this, we use pyramidal (hierarchical) 3D lattice. A priori probability is determined by transition probability from one layer of the lattice to another. It was confirmed that our method improved correct classification rate by about 20% compared with that obtained by the conventional maximum likelihood method or co-occurrence probability method.

Forward modeling represents an emulation technique providing a numerical approximation of a physical process. Forward modeling is often applied to find a solution for problems where no direct inversion techniques are available. Typical examples are the determination of physical quantities that can only be found based on comparisons with successively refined models. Classical forward modeling often relies on learned guesses and practical experience. An overview will be given where this technique can be used profitably as a data analysis tool in remote sensing. We will show typical application areas such as imaging observations of land and water surfaces with optical and microwave instruments and compare the approaches with atmospheric retrieval techniques of a limb sounding instrument. Bridging the gap between practical experience and theoretical concepts, maximum likelihood calculations allow the assessment of forward modeling with respect to the accuracy of information extraction. Information theoretical bounds can be derived from the Cramer-Rao bound and the Kullbach-Liebler divergence. As a result, a maximum likelihood estimation can be defined and be interpreted as fitting the observed likelihood to its true value. This allows a determination of the attainable accuracy of the process to be modelled.

The urban areas represent a vast subject in image interpretation. It is interesting to be able to analyze town development, to make streets maps automatically or just, to mask the urban areas in satellite images. The objective of this study is to extract urban areas from remote sensing images and to make a classification of these areas. The proposed method combines different types of operators. At first, we define automatically a mask of the urban areas by combining a classification algorithm with edge extraction algorithms. Then these urban areas are segmented according to the streets, railways and rivers to obtain urban districts. Finally, we define a measure of urban density. In this paper, we focus on the urban extraction algorithm and the urban segmentation process. This study is performed by IRIT and has been partially funded by CNES agency.

In this paper we present the development of a technique, based on multiresolution wavelet decomposition, for the merging and data fusion of a high-resolution panchromatic image and a low- resolution multispectral image. The method presented here consists of adding the wavelet coefficients of the high- resolution image to the multispectral (low-resolution) data. We have studied several possibilities concluding that the method which produces the best results consists in adding the high order coefficients of the wavelet transform of the panchromatic image to the intensity component (defined as L equals R+G+B/3) of the multispectral image. Using this method, the detail information from both images is preserved. The method is capable of enhancing the spatial quality of the multispectral image while preserving its spectral content to a greater extent. The method presented does not modify the total energy of the multispectral image, since the mean value of each of the added wavelet planes is 0. The method is, thus, an improvement on standard Intensity-Hue-Saturation (IHS or LHS) mergers. We used the method to merge SPOT and LANDSAT (TM) images. The technique presented is clearly better than the IHS and LHS mergers in preserving both spectral and spatial information.

This paper presents a hybrid strategy for the classification of multisource remote sensing images basing on a knowledge representation framework which integrates fuzzy logic and Dempster-Shafer theory and is capable of dealing with possibilistic and credibilistic forms of uncertainty in an unified way. Within the strategy, the salient, innovative aspect here proposed is the use of a novel neural network model for refinement of fuzzy Dempster-Shafer classification rules. The approach has been evaluated by developing real- world applications in the field of water vulnerability assessment and fire risk assessment. Numerical results obtained show that classification benefit from the integration of neural and symbolic frameworks.

For the classification of the radar data, several techniques have been developed, which take the statistical properties of the radar distribution into account and use a priori segmentation to have better contextual information. The introduction of synthetic neo-channels, describing the local texture of radar images, improve the classification process. We also test two different processes to minimize the inter- class confusion caused by the speckle noise: a pixel-by-pixel basis classification which requires a preliminary spatial and/or temporal speckle filtering, or a contextual method without filtering. In the case of the multi-source data classification, we present a fusion algorithm which consists in implementing different statistical rules for radar or optical images.

This paper reports about a quantitative evaluation of pyramid- based schemes performing a feature-based fusion of data from multispectral and panchromatic imaging sensors having different ground resolutions. A critical point is performances evaluation of image data fusion. A set of quantitative parameters has been recently proposed. Both visual quality, regarded as contrast, presence of fine details, and absence of impairments and artifacts (e.g., blur, ringing), and spectral fidelity (i.e., preservation of spectral signatures) are concerned and embodied in the measurements. The aim of the present work is to provide a comprehensive performance comparison on SPOT data among three feature-based schemes for image fusion, as well as on a specific case study on which multisensor observations were available. Out of the three methods compared, respectively based on high-pass filtering (HPF), wavelet transform (WT), and generalized Laplacian pyramid (GLP), the latter two are far more efficient than the former, thus establishing the advantages for data fusion of a formally multiresolution analysis.

We are presenting an algorithm for the detection and tracking of buried linear features under a variety of surface coverages. The buried structures manifest themselves as a few pixels wide bands with contrast and texture changes of the over-ground growth, in high 1 m resolution aerial photographs. Some statistical non-linear filters are used to enhance these features, and their responses is further enhanced by lateral continuity, taking into consideration prior knowledge about the shape of the feature.

Imagery from the new generation very high-resolution sensors, will increase dramatically the geometric scene resolution but it will also decrease the accuracy of the For urban applications in particular, with the spatial properties of the new sensors it will be possible to recognize not only a generic texture window with specific urban characteristics, but also to detect in detail the objects that constitute the 'urban theme.' In this paper a segment based segmentation procedure is presented, based on the gray-scale geodesic morphological transformation and has been successfully utilized to detect built-up objects using only the 5 m spatial resolution panchromatic data of the IRS1-C satellite. The imagery is subsequently classified on a segment basis using a multi-layer perceptron neural network classifier.

We describe a new method for the detection and reconstruction of building in dense urban areas using high resolution aerial images. Our approach begins with the generation of a dense digital elevation model (DEM). A sparse disparity map is densified using a region-based segmentation of the left aerial image: each detected region is tested to be planar in the disparity map. A strategy is proposed to optimize the generation of these planar surfaces taking into account the noise present in the sparse disparity map and the robustness and complexity of different algorithms for planar approximation. The second step of our approach deals with the generation of building hypotheses. Based on the DEM previously computed, geometric and colorimetric criteria are used for the fusion of parallel regions, for the detection of symmetrical regions in the 3D object space and for the reconstruction of roof buildings. Experimental results are presented on a scene in the suburb of Bruxelles with color images at the resolution of 10 cm/pixel.

This paper reports on an experimental study designed for the in-depth investigation of how a supervised neuro-fuzzy classifier evaluates partial membership in land cover classes. The system is based on the Fuzzy Multilayer Perceptron model proposed by Pal and Mitra to which modifications in distance measures adopted for computing gradual membership to fuzzy class are introduced. During the training phase supervised learning is used to assign output class membership to pure training vectors (full membership to one land cover class); the model supports a procedure to automatically compute fuzzy output membership values for mixed training pixels. The classifier has been evaluated by conducting two experiments. The first employed simulated tests images which include pure and mixed pixels of known geometry and radiometry. The second experiment was conducted on a highly complex real scene of the Venice lagoon (Italy) where water and wetland merge into one another, at sub-pixel level. Accuracy of the results produced by the classifier was evaluated and compared using evaluation tools specifically defined and implemented to extend conventional descriptive and analytical statistical estimators to the case of multi-membership in classes. Results obtained demonstrated in the specific context of mixed pixels that the classification benefits from the integration of neural and fuzzy techniques.

Optimized combination, regularization, and pruning is proposed for the Parallel Consensual Neural Networks (PC-NNs) which is a neural network architecture based on the consensus of a collection of stage neural networks trained on the same input data with different representations. Here, a regularization scheme is presented for the PCNN and in training a regularized cost function is minimized. The use of this regularization scheme in conjunction with Optimal Brain Damage pruning is suggested both to optimize the architecture of the individual stage networks and to avoid overfitting. Experiments are conducted on a multisource remote sensing and geographic data set consisting of six data source. The results obtained by the proposed version of PCNN are compared to other classification approaches such as the original PCNN, single stage neural networks and statistical classifiers. In comparison to the originally proposed PCNNs, the use of pruning and regularization not only produces simpler PCNNs but also gives higher classification accuracies. In particular, using the proposed approach, a neural network based non-linear combination scheme, for the individual stages in the PCNN, produces excellent overall classification accuracies for both training and test data.

This paper describes a novel recursive and unsupervised learning method for extracting information from remote sensing images. Usually, the amount of data on these images is large, and the number of mixed pixels is important. Therefore, an unsupervised learning or clustering can be useful in the analysis of these data. An unsupervised neural network algorithm is used for initial segmentation of the spectral data space of remote sensing images. To discover concepts, a recursive region aggregation method is proposed. This method has been tested and validated with several remote sensing images. An urban zone image is used to illustrate this learning method which provides a way for fast and automatic segmentation of remote sensing images. In order to improve the efficiency of concept extraction some spatial information is incorporated into the aggregation procedure.

A supervised technique for training Radial Basis Function (RBF) neural classifiers is proposed. Such a technique, unlike traditional ones, considers the class-memberships of training samples to select the centers and widths of the kernel functions associated with the hidden neurons of an RBF network. The proposed method has significant advantages over traditional ones in terms of classification accuracy and stability of the network. Experimental results, carried out on a multisensor remote-sensing data set, confirm the validity of the proposed technique.

We propose in this paper an integration method of the radar information in multispectral images without disturbing the spectral content. The main problem is to define a fusion rule that allows to take into account the characteristics of these images. Also, the main purpose of this paper lies in defining a new fusion rule performed in the redundant wavelet domain. This rule is based on the Mahalanobis distance applied on the wavelet coefficients. Instead of comparing coefficient-to- coefficient, the distance-to-distance comparison is performed. In this case the selected coefficient in the fused image will be the one that presents the large distance. This approach is applied to fusing the infrared band of SPOT with, respectively, RADARSAT and ERS images. The results show that spectral information is well preserved and there is a better information on the texture and the area roughness.

In this work we present a new algorithm to generate the texture print of a region in an image. For this texture analysis, a texture print is found by means of counting the number of changes in the sign of the derivative in the gray level intensity function by rows and by columns, over a region with size N X N. These two histograms are represented as a unique string R of symbols. Therefore, a string-to-string correction problem as placement rules of elements (primitives) obtained statistically is used. In order to discriminate different texture regions a distance measure on strings based on minimum-cost sequences of edit operations is computed, this measure is the Leveshtein distance. The proposed algorithm is useful to discriminate between urban areas and rural areas due to the change in their textural aspect.

The context of this study is the 3D reconstruction of urban scenes from aerial images. We intend to estimate a dense depth map precise enough to be exploited by recognition algorithms. In this paper, we show how a multi-view approach made up of very simple and automatic operations can achieve this goal. Unlike 2-view stereovision methods, we do not exploit a disparity map for depth estimation. The proposed method consists in directly scanning depth. For each depth hypothesis, a reference image is projected by using a planar perspective transformation. The correct hypothesis is found pixel by pixel by minimizing a simple matching criterion based on a gray level comparison. We calculate this criterion for each pixel of the reference image, each depth hypothesis and each baseline formed with the reference image and and an other image of the sequence. The estimated depth map obtained with a synthetic aerial image sequence, shows that exploiting several images with different baselines reduces the reconstruction errors due to noise and false matches. We have implemented an algorithm composed of simple automatic computations, that should be highly parallelizable.

Many image compression techniques have been developed for remote sensing imagery over the last thirty years. What are considered as standard techniques such as the use of principal component analysis, discrete cosine transform, predictive coding, etc. have shown their limitations. Wavelet transform techniques have been increasingly used in recent years. In this paper a new and efficient technique is presented that provides a nearly lossless compression of the multichannel remote sensing imagery by combining the use of wavelet decomposition, non-uniform quantization, arithmetic coding, and geometric vector quantizer (GVQ) to achieve the compression task with very minimal loss. The detailed procedures will be illustrated with real remote sensing images.

Spatial details of surfaces acquired by means of imaging spectrometers and multiband cameras are degraded by many factors. The atmosphere placed between the instrument and the surface, optical aberrations and tracking errors are some sources. Due to these causes, the photons coming from the instantaneous field of view pertaining a certain pixel, are spread over a larger number of picture elements, causing a spatial filtering of the image. Natural surfaces are rarely composed of a single uniform material and, therefore, blurring causes also a mixing of spectra of mineralogic different units on the surface. The problem of image sharpening is then linked to that of spectral unmixing. In this work, we compare the use of different statistical techniques, as Principal Component Analysis, Linear Spectral Unmixing and Spectral Clustering for image sharpening purposes.

Radiative transfer models have been the most popular approach to patten recognition from remotely sensed images of the Earth. Unfortunately, these methods have several limitations regarding the Lambertian assumption on natural surfaces, limited semi-empirical design, restricted number of spectral channels actually used, and inaccessibility of ground truth data. Multi-spectral techniques based on the application of unsupervised neural networks could substantially reduce the complexity of satellite image analysis and contribute to the improvement in pattern classification. The current work is concerned with ocean property extraction from the ADEOS OCTS sensor using a hierarchy of multi-spectral processing involving self-organizing feature maps and expert rules on water suspended substance concentrations. After the initial image correction for the 'standard atmosphere,' water pixels are separated and classified into a large number of water colors. This paper presents the final stage of the analysis defining concentrations of phytoplankton and other water suspended substances within image pixels. As spectral radiances are non-linearly depended on the backscattering and absorption, the algorithm is only based on widely known absorption distributions in the visible spectrum and expert rules considering pixel temperature, atmospheric contamination, and proximity to land and clouds. The method was tested on nine OCTS scenes portraying coastal sites around the Pacific Ocean.

In the history of SAR image processing, many algorithms have been proposed to tackle the problems of segmentation, classification and edge detection. They are typically heuristic in basis, and more successful on some types of imagery than others. With the development of global optimization methods it has now become possible to produce optimal techniques; that is, those which can genuinely achieve the optimal solution of the posed problem. The problem is characterized by an objective function and the chosen optimization technique. The most successful and wide-spread method has been simulated annealing and we detail its application in the fields of segmentation and classification. In particular, we detail how to optimally quantify the relationship between competing terms within the objective function. The performance of the resulting algorithm on various SAR imagery is given.

The recent advance and diffusion of airborne and satellite SAR systems makes the use of multitemporal SAR images of practical interest for monitoring and control applications, especially when aiming at the identification of moving objects. Change detection is a powerful technique, which allows the detection of slowly moving targets. This is obtained by subtracting on a pixel-by-pixel basis the intensity of two SAR images collected at different times and comparing the absolute value of the difference to a detection threshold. Under ideal conditions, the fixed background is totally correlated and cancels out completely. On the contrary, a slowly moving target changes its position in the two images and is not cancelled. Therefore only its echo crosses the threshold. In practice, a number of factors cause the deviation from the ideal conditions; among the others the temporal decorrelation of the scene due to the internal clutter motion and the presence of misregistration and miscalibration errors. The prediction of change detection performance is essential to the design of SAR based systems for the moving target detection and should take into account all of the possible mismatch causes. In the present paper we aim at a complete mathematical characterization of the performance of change detection under non-ideal conditions. Specifically, Section 2 introduces the change detection technique and discusses some of the possible deviations from the ideal conditions. The analytical derivations are presented in Section 3, together with the discussion of the achieved results. To validate the theory, the obtained performance prediction is compared to the results obtained with real SAR images in Section 4.

Remote sensing is a domain where one of the biggest important problems is the interpretation of large-sized images. Thereby, it is not possible for experts to analyze the ceaseless image streams. In practice, there is a growing interest in understanding concepts discovered in classified images. Our approach to image classifications is based on the conceptual clustering algorithm, Cobweb and its extensions. In general, these algorithms produce tree-structured clusters. However, once the hierarchies are built, the remote sensing experts need to compare and to synthesize the obtained hierarchies in terms of conceptual similarities. Two algorithms are described which produce a synthesis of hierarchies. The first algorithm can be used to synthesize results generated by heterogenous hierarchical classifiers, such as K-means, Unimem, Labyrinth. The second algorithm is an extended version of Cobweb. The experiments carried on urban zones have shown the universality and the efficiency of our approaches.

Rules for extracting objects and features from remotely sensed data tend to become case specific and thus lack of generalizability beyond the training area. To alleviate the severity of this problem, we propose to low fidelity space- based imagery to extract objects in the context of multisensor fusion. The test site is Sarajevo, and the data sets are LANDSAT TM multispectral and Canadian RADARSAT synthetic aperture radar (SAR) data. The software environment is IMaG system developed by Susquehanna Resources and Environment, Inc. Since IMaG allows one to perform spectral and spatial integration using a scripted programming language, objects existing in two dissimilar sensor domains can be merged and extracted by using soft decision rules that are more generalizable than hard decision rules based on conventional supervised classification methods. Objects extracted in the test site include the built-up area, the runway, rivers, pine forests, and so on.

Automatic analysis of remote sensing images faces different problems: context diversity, complexity of information. To simplify identification and to limit the search space, we use extra data and knowledge to help the scene understanding. Diversity and imprecision of information sources generate new problems. The fuzzy logic theory is used to solve the problem of imprecision. Many extraction algorithms are used to provide a more reliable result. Extraction may be performed either globally on the whole image or locally using information of data bases. Each extractor produces a map of certainty factors for a given type of geographic features according to their characteristics: radiometry, color, linear, etc. Maps contain wrong detections due to imperfections of the detectors or non- completeness of generic models. So, we generate a new map using fusion to have a best credibility used to compute a dynamic programming. It finds an optimal path even if the linear feature is partially occluded. But the path is generally erratic due to noise. Then a snake-like technique smooth the path to clean the erratic parts and to tune the level of detail required to represent the geographic features on a map of a given scale. The result is used to update data bases.

This paper presents the characterization of Land Surface Temperature (LST) variability according to land cover, in order to derive the properties of evapotranspiration and improve the monitoring of a catchment. The land cover can be represented by its Normalized Difference Vegetation Index (NDVI) and first results underscore the relation between T and NDVI at NOAA-AVHRR pixels scale. However, due to their rough resolution, these pixels include several land cover types and this study revealed not useful for catchment monitoring. Therefore, Land Surface Temperature has to be specified with a more precise representation. We employ a physical model of temperature, which requires several parameters such as proportion and emissivity for each component within the pixel; these values are obtained with learning process using high resolution data such as Landsat TM. These results are then extrapolated to the global region with NOAA-AVHRR acquisitions and allow to analyze the land cover effects on Land Surface Temperature variability. By this way, the characterization of evapotranspiration according to land use for a global catchment is improved.

This article addresses the application of pattern recognition techniques for the analysis of urban and rural areas based on high resolution panchromatic space-borne imagery. In particular, we focus on the problem of finding stable features that allow one to distinguish between urban, industry, and vegetation, categories that are important for understanding the urban pressure on rural areas. The novelty of the paper is the detailed and thorough analysis of the usability of more than 30 texture features in this type of analysis. It is concluded that for the application and the data at hand the homogeneity and contrast features derived from the gray level co-occurrence matrix are the most successful for SPOT panchromatic data.

In this paper, we shall describe a new method for contextual classification of remotely sensed images. A compound deterministic pseudo annealing-Markov random field model that estimates the correct label for each pixel is suggested. The contextual information is used via the developed context function that depends on the probability function for the adjacent class labels. This function is then formulated as a discrete Markov Random Field (MRF). The global energy function to be minimized is made up of the adaptive a priori probability of classes and the context function. Without using the Metropolis criterion, the optimization procedure consists of selecting the new configuration that corresponds to the minimal value of the global energy function. We call such a procedure of optimization a Deterministic Pseudo Annealing (DPA). The method has been tested and evaluated on real multispectral image provided by the SPOT satellite. The results obtained have the same, or nearly the same, accuracy as those obtained with simulated annealing (SA)-based method and Iterated Conditional Modes (ICM)-based method. The convergence of the proposed DPA approach is better than SA method and very close to ICM method.

The impact of natural and man-made disasters on the social and economic progress is going to become more significant, making necessary to consider natural disasters reduction. Therefore civil protection and resource managers need elements to make quicker and better decisions on a day-to-day basis, so giving the start to an emerging world-wide remote sensing market. A deep analysis on the potential users, mainly devoted to Mediterranean basin, highlights that existing and/or planned systems are not able to completely satisfy their requirements. To fulfill this gap, Italy decided to promote the SkyMed/COSMO system, presently financed by the Italian Space Agency. SkyMed/COSMO is a constellation of small satellites for observation, remote sensing and data exploitation for risks management and coastal zone monitoring, conceived to provide products, services and logistics to both institutional and commercial remote sensing users on global scale. Furthermore the system is able to satisfy a broad spectrum of important applications also in the field of the resource management, land use and law enforcement. The SkyMed/COSMO current system architecture foresees a constellation of small satellites in two different orbit planes composed by 4 satellite equipped with X-band SAR and 3 satellites equipped with optical sensors. The system is characterized by good spatial resolution, day and night/all-weather imaging capability and by a very good revisit time. The program, currently in phase B, is carried out by an industrial consortium lead by Alenia Aerospazio.