A new automated registration method using disparity detection is proposed. In this method, a new pair of control points is automatically generated on a point which has the largest disparity between images to be registered. The procedure is repeatedly applied to the images until the maximum disparity falls below a threshold. Considering the images to be registered as sequential shots of a video image, the disparity is detected as optical flow vectors. The method enables us to ensure the accuracy of registration.

The topic of this contribution is a matching between property borders of cadastral maps and field borders identified in satellite imagery (Landsat TM). For reasons of efficiency, the search space for the matching is spatially limited. Sufficient contextual information in a small spatial environment, therefore, is of paramount importance. The short edges of lots of long and narrow fields cannot be extracted in the same go as the well defined long edges because they are perceived edges rather than real step edges. A two step algorithm is needed for their extraction. Perceptual and conventional edges can be combined to supply the necessary local contextual information for a robust and efficient matching to take place. The matching algorithm is a refined version of probabilistic relaxation.

The main feature of this paper is to show that the key point of two different problems tackled by neural approaches -- pairing pattern and function approximation -- lies in the choice of the regularization term in the function which is minimized by the neural approach. After the description of a new algorithm allowing the matching between two set of points with a nonuniform distribution in the plane, and a registration based on the regularization theory, we show that a multitemporal analysis can easily be done.

The algorithm for search of the correspondence between the points from two planar finite point sets is presented. The protective transformation between the point sets is supposed, but points without correspondence in the other set can be presented in both sets. The algorithm is based on the comparison of two projective and permutation invariants of five-tuples of the points. The most hopeful five-tuples are then used for the computation of the projective transformation and that with the maximum of corresponding points is used.

In the framework of an image interpretation system for automatic cartography based on remote sensing image classification improved by a photo interpreter knowledge, we developed a system based on neural networks which simultaneously produce fuzzy rules, with their linguistic approximation as well as final classification. This paper describes the succession of steps used with this aim in view. Particularly it investigates the application of mutual information criteria to simplify fuzzy rules.

A multistep method for segmentation of feature space using triplet decision tree is developed, and another approach to cope with uncertain samples by extended Bayesian discriminant function is introduced. The latter has the lower limit for posterior probability of classification. The triplet-decision tree includes a division-wait mechanism that postpone the decision about uncertain samples which are in marginal area and not able to be classified to any categories definitely. The third node is generated for such samples. Improvement of the triplet tree method is made by introducing linearly-combined variables related to principal components. Flexible and effective segmentation is accomplished by this refinement. Results of experiments by simulation data and real remotely-sensed data are compared by the two methods in the viewpoint of cutting of feature space and classification accuracy. When the normality or representability of sample is hold, classifier with extended quadratic discrimination function has the best performance. The advantage of triplet tree appears when categories are diversified in nature or training samples have poor representabilities.

The SAIL model, a canopy reflectance model, was used to simulate narrow-band reflectance of overstory/background compositions to study the effect of the background on the estimation of the coniferous forest LAI based on remotely sensed data. We have simulated several mixed targets with a pine tree canopy and different backgrounds, including understory vegetation, soil and litter. For each type of mixed target we have modelled the reflectance for several LAI. The modeled data were used to evaluate the performance of the broad and narrow band NDVI for predicting the LAI. Results show that, for low LAI, the type of background contributes strongly to the reflectance of the mixed targets. Furthermore, the way the understory affects the mixed signal depends significantly on the vegetation species. The sensitivity of the NDVI for estimating the pine canopy LAI depends on the type of background and it was verified that mixed targets with non-vegetation backgrounds have larger sensitivity than the ones with vegetative backgrounds. The results show that the NDVI, calculated with broad or narrow bands, is not adequate to predict the LAI of open pine stands, when one does not known the type of background that is underneath the pine canopy.

Several measures assessing the accuracy of land-cover classification are available, e.g., overall and class- averaged accuracies. Also kappa statistic is widely used for this purpose. In this article, we discuss the properties of these criteria, and point out that the kappa statistic has an unfavorable feature. We propose alternative coefficients based on Kullback-Leibler information. Further, significance tests for the difference between the coefficients derived by classification results are established.

The k-NN rules and their modifications offer usually very good performance. The main disadvantage of the k-NN rules is the necessity of keeping the reference set (i.e. training set) in the computer memory. Numerous algorithms for the reference set reduction have been already created. They concern the 1-NN rule and are based on the consistency idea. The 1-NN rule operating with a consistent reduced set classifies correctly, by virtue of consistency, all objects from the original reference set. Quite different approach, based on partitioning of the reference set into some subsets, was proposed earlier by the present authors. The gravity centers of the subsets form the reduced reference set. The paper compares the effectiveness of the two approaches mentioned above. Ten experiments with real data concerning remote sensing data are presented to show the superiority of the approach based on the reference set partitioning idea.

Synthetic aperture radar data may contain useful information about small structures like rivers and man made structures. Using the Markov random field based segmentation algorithms, that perform quite well on homogeneously textured areas (i.e., agriculture land cover), these structures may be lost if they are small (1-2 pixel wide). The merging of various sources of information at a low level of the Markov random field region label process, makes it possible to recover at least partly the fine structures in the SAR data. The data fusion makes use of Bayesian inference about the Markovian property of neighborhood systems. This article shows that the proposed method is valid and technically feasible, based on extensive validations.

The study of vegetation repartition and evolution is a wide research field and application task for environmental modeling. The objective of our study is to discriminate different vegetation types by their temporal evolution. For that purpose, we use two different sensors: the SPOT sensor provides monthly data with a sufficient spatial resolution, while the NOAA sensor provides daily data, but with a poor spatial resolution. Combining these two complementary sensors seems to be a promising way to lead the study. We propose here a three-step experiment showing that the simultaneous use of these two sensors allows us to obtain a fine segmentation of land cover.

The study of urban area is an important problem in image interpretation. It is interesting to be able to analyze town development on satellite images or to mask urban areas. The objective of this study is to extract urban areas from remote sensing images and to make a classification of these areas. The detection algorithm combines different types of operators in order to improve the final detection. We separate urban areas from the other type of regions (vegetation, rivers . . .). Then the urban areas are classified into various under-classes (dense urban areas, suburbs . . .). This study has been performed by IRIT in the frame of the CNES program on studies and research on automatic analysis and interpretation of SPOT images.

In this paper, a novel approach is proposed, which allows for an efficient reduction of the amount of visual data required for representing structural information in the image. This is a multistage architecture which investigates partial correlations between structural image components. Mathematical description of the multistage hierarchical processing is provided, together with the network architecture. Initially the image is partitioned to be processed in parallel channels. In each channel, the structural components are transformed and subsequently separated, depending on their structural significance, to be then combined with the components from other channels for further processing. The output result is represented as a pattern vector, whose components are computed one at a time to allow the quickest possible response. The input gray- scale image is transformed before the processing begins, so that each pixel contains information about the spatial structure of its neighborhood. The most correlated information is extracted first, making the algorithm tolerant to minor structural changes.

A new approach to the detection of above-ground from a pair of stereoscopic images in a general urban context is proposed. It includes a stereoscopic matching stage well- adapted to our task in order to provide a digital surface model (DSM). Then a segmentation of the DSM is performed, and regions are classified as ground or above-ground. The interest of the method is its ability to manage extended above-ground with several heights and any shape, as well as the case of a sloping ground. An application to digital terrain models (DTM) generation in urban areas is discussed. Assessment of both above-ground extraction and DTM generation on difficult scenes shows the feasibility of the approach.

We address in this paper the problem of defining fuzzy relationships between objects for pattern recognition purposes. We distinguish two kinds of spatial relationships. Some of them are well defined if the objects are crisp, like adjacency, inclusion or other set relationships. But since they are highly sensitive to errors or imprecision in segmentation, more useful measures can be obtained by fuzzifying these concepts. We define such measures using fuzzification principles or direct translation of binary equations into fuzzy ones. Other relationships are inherently vague concepts, like relative position. Fuzzy definitions of such relationships are then more consistent than crisp ones. We propose definitions of relative position between crisp or fuzzy objects, and illustrate them on buildings in aerial images.

We present a new method for perceptual grouping of pixels into roads after crest lines detection in satellite pictures. First the visual properties expected from the groupings are modelled as a quality function similar to active contour functions. They involve curvature, gray levels and co-circularity. This function is computed recursively and optimized from a local to global level with an algorithm related to dynamic programming. The final groupings are then selected according to their global quality. Applied to satellite images, the method proved its adaptability and its robustness to noisy environments. The results showed how the use of visual properties can provide an effective segmentation with no prior knowledge of the scene. This segmentation can be used to initialize a high level interpretation process or give a first description of the scene to an interactive decision system.

Applications of new texture features in SPOT image segmentation are presented in this paper. The set of texture features is based on morphological residues of opening and closing by reconstruction. In texture classification, this set of features is proven much more robust to noise than feature set derived from traditional morphological residues. An optimization algorithm is established to search for the optimum feature subset, and a minimization of window size is evaluated to obtain better classification accuracy. In experiments of various noise circumstances, it is found that this feature set bears quite high texture classification accuracy compared to other texture classification methods. In application of SPOT image segmentation by texture classification, an optimal feature subset with the supervised Gaussian maximum likelihood classifier is employed. To improve the segmentation performances, post- processing is added. Comparisons with other segmentation methods are made.

The definition of macro-texture proposed here comes from granulometric analysis which is a way to describe statistically the size of the connected components in a binary. The pixels of the binary image are classified according to the quantitative descriptors provided by this analysis. The whole procedure is applied here to the mapping of the density of the vegetation cover in arid lands from SPOT and thematic mapper.

The multilook polarimetric maximum likelihood classifier based on the Wishart distribution supposes no variation of the backscattering of the underlying scene. For clutters verifying the 'product model', we here present the use of a K-distribution and compare this classifier to the one based on the Wishart distribution. A simple way to obtain a full polarimetric filter by filtering a set of adequate powers is also given. We show how filtering and segmentation of the raw data improve the classification results.

Imaging spectrometers acquire images in many narrow spectral bands but have limited spatial resolution. Spectral mixture analysis (SMA) is used to determine the fractions of the ground cover categories (the end-members) present in each pixel. In this paper a new iterative SMA method is presented and tested using a 30 band MAIS image. The time needed for each iteration is independent of the number of bands, thus the method can be used for spectrometers with a large number of bands. Further a new method, based on K-means clustering, for obtaining endmembers from image data is described and compared with existing methods. Using the developed methods the available MAIS image was analyzed using 2 to 6 endmembers.

In this paper an algorithm to detect changes in multispectral and multitemporal remote-sensing images is presented. Such an algorithm makes it possible to reduce the effects of 'registration noise' on the accuracy of change detection. In addition, it can be used to reduce the number of typologies of detected changes in order to better point out the changes under investigation.

The purpose of this paper is to define the mass functions for a set of multiple mixed hypothesis in an context of Dempster/Shafer (DS) theory which offers an interesting tool to combine data providing from heterogeneous sources more or less reliable by managing imprecision and uncertainty. This is particularly important when dealing with multi-modality imaging (satellite image), where the fusion of information increases the global knowledge about the phenomenon while decreasing the imprecision and uncertainty about it. This theory also enables us to assign masses to 2^D elements (D: decision space) rather than to D elements as in probabilistic theory. The DS has been used in many applications in the field of image analysis, but without its all powerful. When using with only simple hypothesis (an object belongs to only one class), the theory falls in the probabilistic case, which is considered as a particular case. Bloch and Barnett attempt to use double hypothesis but their method still remains particular and restrictive. We propose in this paper a method to extract for a class the consonance and dissonance degrees among several classifiers (methods), and the integration of these terms to initialize the mass functions with multiple mixed hypothesis in order to use the orthogonal Dempster/Shafer Rule. The problem must be viewed from multiclass, multi-sources (images) and multi- point of view (methods or classifiers used) context. We first show how our method works with 1 -- image, 2 -- classifiers, and 2 -- hypothesis and then generalize for P - - images, K -- sources and 2^D -- hypothesis.

An innovative perspective to analyze the quality of digital elevation models (DEM) it is presented here. It is based on the fractal Brownian model which is suitable for the description of natural surfaces. Applications to across- track interferometric synthetic aperture radar (SAR) DEM and classical photogrammetric DEM is here accomplished showing the viability and usefulness of the proposed method.

Though accurate motion compensation for SAR satellite imaging is not an acute problem at present, it will probably become one in the near future with increase in the ground resolution of available sensors. This problem is already nowadays extremely pertinent for airborne SAR imaging, especially with slow and/or unmanned aircrafts. Image registration is often the first and highly critical step in many operational uses of SAR imaging. We illustrate some of those applications with highly distorted airborne SAR images. Those applications are of two main classes, one being the registration of two similar SAR images without an external reference, and the second being the registration of a SAR image with another type of geographical data such as an optical image or map. The local similarity of the images in the first class of problems leads us to adapt a technique from optical flow motion evaluation. We used a local correlation technique with a low dimensional parametrization of the motion to be estimated. The second class of problems is dealt by first extracting linear (sometimes pointwise or areal) features and then matching them by means of a multi hypothesis filtering. We shall describe here four applications of image registration: (1) The first one is the registration of single-look images obtained within the same overflight in order to compensate for trajectory errors not detected by the aircraft navigation system. This registration not only allows the multi-look incoherent summation, but also the refocusing of the single-look images to increase azimuth resolution. This process is the type of 'auto-focusing' known as frame-drift. (2) Multi-look registration also provides an accurate evaluation of ground speed of the carrier aircraft that can be exploited for internal navigation unit hybridization, even in the absence of cartographic knowledge of the area flown over. This application is called 'odometering.' (3) Multi-look registration is furthermore extremely critical for moving vehicles detection and their proper motion estimation using multi-look imaging on wide beam antennas. This is one version of the problem known as 'moving target indication.' (4) Registration of a SAR image with a map or an optical image of a given area is also used for accurate localization of objects detected in geographical coordinates, for their designation to other systems. It may also be used for accurately localizing the carrier air-vehicle, especially when unmanned. This is the 'navigation resetting problem.'2955

Median filter (MF) is a powerful tool for impulsive noise removal in digital signals and images: noisy samples do not affect its output, but are discarded as outliers. The basic scheme of median filter has been specialized to remove noisy spikes with little distortion, that is without modifying noise-free pixels, like the rank-conditioned median (RCM) filter, recently introduced by the authors. In this work a spatially adaptive RCM scheme (ARCM) is proposed with the aim at extending its filtering capability also to additive and multiplicative noise models. Accordingly, only pixels having botndary ranks are adaptively replaced with the sample median, while the others are left unaltered: pixel replacements are conditioned to their ranks, i.e., positions of their values after sorting in ascending order, based on an estimate of local SNR and on the response of a simple detector of structured edges. An adaptivity function is empirically designed, and a unified framework is developed to deal with both additive and multiplicative noise models. Results are presented on images corrupted with several noise models, as well as on true Synthetic Aperture Radar (SAR) images affected by speckle noise. Keywords: Image processing, Edge-preserving smoothing, Impulsive noise, Median filter, Rank filters, Adaptive filtering, MMSE filtering, Multiplicative noise, Synthetic Aperture Radar images, Speckle.

In this paper we show that a Line Processor can solve most of problems given by stereovision. Our architecture supports algorithms that often reach real-time execution. Several approaches, according to the image texture have been considered. This architecture is well suited for on board systems.

With the wavelet transform theory and the Markov random model, this paper presents an unsupervised multiresohition segmentation method to segment the textured SAR image. This method specially includes a step to estimate both the optimal number of texture classes and their model parameters without supervision. In order to interpret the results of the unsupervised segmentation as well as to understand the whole polarimetric SAR image, this paper also develops an interpretation approach which jointly utilizes the target decomposition theory and the identification technique of the scattering mechanism. Experimental results are presented for demonstration.

Novel robust algorithms using combined rank and M-estimations applicable to image and signal processing are proposed. They are based on analysis of statistical characteristics of images and permit to enhance the quality of radar images by means of effective speckle suppression, impulsive noise removal and edge preservation. The advantages of these digital algorithms are demonstrated by simulated images and real data.

The aim of the present analysis is the design of a reliable automatic updating process for the IGN BDCarto. Geographic evo— lution is a dynamic phenomenon, whereas cartographic updating, more static, requires the knowledge of a geographic state at one moment. In the GIS context, digital data gets a dynamic dimension, and the updated database must be in coherence with the initial one. Thus, two aspects have to be focused on : detect the changes and keep the inter-temporal coherence. Tests, composing an automatic updating in a test-area with an existing process, showed not reliable enough results. The main problems we met with such a process is the variability of interpretation, which drowns the real changes supposed to be detected. The difficulty is reinforced by the generalization operations imposed by the data-base specifications. To minimize errors induced by those aspects, an improved updating process may be envisaged, in which knowledge of a priori evolutions would be introduced. The proposed solution consists then in including constraints, probabilistic for the evolutions and geometric for the generalization, in a fuzzy classification.

How to calculate Legendre moments of images by less computation is a very important problem for the application of orthogonal moments in pattern recognition and image processing. in the present paper we propose the fast calculation to characterise 1-D signals and 2-D images by Legendre moments. We present at first the projections of 1-D signals onto Legendre polynomials and deduce their recursive calculation. We then introduce the scaled Legendre polynomials, the Legendre moments of 1-D signals with arbitrary window size and theirfast calulation. The implementation in discrete cases is presented and its computational complexity is analysed as well. To apply our method to 2-D image processing and recognition, the recursive algorithm is generalized to 2-D cases. With the approach presented, the computational complexity to characterise images by Legendre moments is greatly reduced and the implementation is simple. As the moments are widely used in image filtering, image segmentation, texture analysis and pattern recognition, the use of the algorithm proposed could greatly reduce the computational complexity of such tasks.

There has been a great deal of interest recently in pattern recognition and classification for remote sensing, using both classical statistics and artificial neural networks. An interesting neural network is Kohonen's seif-organising map (SOM), which is a clustering algorithm based on competitive learning. We have found that seif-organisation is a neural network paradigm that is especially suited to remote sensing applications, because of its power and accuracy, its conceptual simplicity and efficiency during learning. A disadvantage of the Kohonen SOM is that there is no inherent partitioning. We have investigated a natural extension of the SOM to multiple seif-organising maps, which we call MSOM, as a means of providing a framework for various remote sensing classification requirements. These include both supervised and unsupervised classification, high dimensional data analysis, multisource data fusion, spatial analysis and combined spatial and temporal classification.