Through the EU FP7 Horizon 2020 ECOPOTENTIAL project, a novel and innovative approach to classification has been developed, which is termed the Earth Observation Data for Ecosystem Monitoring (EODESM), and has been built on concepts behind an implementation of the Earth Observation Data for Habitat Monitoring (EODHaM) system generated as part of the EU FP7 BIOSOS project, applied to Very High Resolution (VHR) Worldview data. The EODESM system facilitates routine classification of land covers according to the Food and Agricultural Organisations Land Cover Classification System (FAO-LCCS), translates these to other taxonomies (including General Habitat Classifications; GHCs) and facilitates routine detection of change and the generation of maps indicating the causes and consequences of such change.

A first attempt to adopt an automated selection of optimal processing local scales within the LG- Mod algorithm is presented. Aimed at Sea Surface Wind (SSW) directions retrieval from Synthetic Aperture Radar (SAR) data, the novel LG-Mod multi-scale analysis mainly relies on an accuracy maximization criterion, based on the use of the marginal error parameter. Results achieved by the algorithm on both artificial and real SAR images appear very encouraging.

A crucial challenge among remote sensing communities is the availability of open data. Several European and international programs address this aim offering a useful support to many applications. The detection of changes in ecosystems for the monitoring of biodiversity is one of the main goal for the commitments undertaken not only by the European Habitats Directive (92/43/EEC) and the Convention of Biological Diversity (CBD) but also the UN Sustainable Development Goals (SDG). This work aims to show the usefulness of free available data for the detecting of changes in (semi-) natural grasslands ecosystem for three different test site in Italy and Portugal. Due to its request for only one multispectral remote sensed image at time T2 and a land cover map at time T1, the Cross Correlation Analysis (CCA) algorithm was preferred for the detection of changes from (semi-) natural grasslands to other land cover. A Sentinel-2 image as T2 and the (semi-) natural grasslands Copernicus layer as T1 were considered. The accuracy assessment of the change maps was based on stratified random sampling weighting the Overall Accuracy (OA) by the proportions of the study area represented by the map classes. The reported finding for OA % of 94.21±0.10, is encouraging and promising as a support to policy makers.

The aim of this paper is the development of an algorithm, based on expert knowledge, for the Land cover classification of an Italian Grassland Area. To accomplish this task, a dataset composed by 4 Worldiew-2 (WV-2) images, at 2 m of spatial resolution, has been considered. Despite their poor spectral resolution, Very High spatial Resolution (VHR) data allow the identification of individual objects by means of the information in the relationship between adjacent pixels, including texture and shape. For this reason a Geographic Object-Based Image Analysis (GEOBIA) approach consisting in a rule set based on the elicitation of expert rules concerning phenology, spatial features and agricultural practices in conjugation with prior spectral knowledge, has been used The study area, of almost 500 kmq, is located in Southern Italy (Puglia Region) within the Natura 2000 “Alta Murgia” site (SCI/SPA IT9120007, according to Habitat Directive 92/43 and Bird Directive 147/2009), partly designated as a National Park as from 2004. Semi-natural dry grasslands cover almost 24% of the total area of the site which represent one of the most important areas for the conservation of this kind of ecosystems in Europe.

This paper focuses on the application of a semi-automatic unsupervised change detection algorithm called Cross Correlation Analysis (CCA) to the detection of (semi-) natural grasslands changes at Very High Resolution (VHR). A reference validated Land Cover/Land Use map at time T1 and only one satellite image at time T2, with T2>T1, are required to detect changes occurred at T2 in the selected target class. This approach offers the possibility to reduce the costs of change detection when the acquisition of multi-seasonal VHR images at time T2 for supervised change detection is too expensive or when no archive VHR image is available in the past for unsupervised comparison between T1 and T2 images. A summer Worldview-2 image for a Natura 2000 test site was considered and the results appear encouraging.

The derivation of habitat maps is enhanced if land cover maps are used as basis for the mapping procedure. In this study, a supervised learning framework is proposed to perform object-based classification to General Habitat Categories. A Land Cover Classification System map is used as basis, and an approach to generate numerical features from the object land cover class names and attributes is introduced. An additional number of spectral, morphological, and topological features are extracted from very high resolution satellite imagery and classification accuracies up to 80.4% for 14 classes are reached. Inclusion of LiDAR (Light Detection And Ranging) data or proposed texture analysis features, improve accuracies to 86% and around 83%, respectively, with the latter proving as promising surrogates of LiDAR data features. The method outperformed rule-based approaches, indicating its potential in accurate and labor- and time-efficient habitat classification.

Earth Observation (EO) images have been extensively used to provide a synoptic view of land cover/use (LC/LU) patterns and land cover/use changes. Land covers are not as clearly relatable to biodiversity in comparison to habitat classifications which can provide more scope for biodiversity monitoring. The main purpose of the paper is to provide an automatic general framework for translating LC maps (in LCCS taxonomy) into habitat maps (in GHC taxonomy) by means of VHR remote sensing data.

This paper presents a method for hedgerow extraction from very high resolution optical images using image segmentation. The method is based on a connected component region growing approach followed by an object based image analysis filtering allowing to introduce high level knowledge about the characteristics of the desired hedgerows. The proposed approach is used as a part of the EODHaM image processing system developed in the framework of the BIO SOS FP7 project.

To date, the automatic or semiautomatic transformation of huge amounts of multisource multiresolution spaceborne imagery into information still remains far below reasonable expectations. The original contribution of this paper to existing knowledge on the development of operational automatic remote sensing image understanding systems (RS-IUSs) is fourfold. First, existing RS-IUS architectures are critically revised. In this review section, the two-stage stratified hierarchical RS-IUS model, originally proposed by Shackelford and Davis, is identified as a subclass of the parent class of multiagent hybrid systems for RS image understanding, which is potentially superior to the two-stage segment-based RS-IUS architecture that is currently considered the state-of-the-art in commercial RS image-processing software toolboxes. Second, this paper highlights the degree of novelty of an operational automatic near-real-time well-posed model-driven application-independent per-pixel Landsat-like spectral-rule-based decision-tree classifier (LSRC) recently presented in RS literature. Third, five original downscaled implementations of the LSRC system are proposed to be input with a multispectral image whose spectral resolution overlaps with, but is inferior to, Landsat's. These five downscaled LSRC implementations are identified as the Satellite Pour l'Observation de la Terre-like SRC, the Advanced Very High Resolution Radiometer-like SRC, the Advanced Along-Track Scanning Radiometer-like SRC, the IKONOS-like SRC, and the Disaster Monitoring Constellation-like SRC, respectively. LSRC, together with its five downscaled implementations, called the integrated SRC system of systems, is eligible for use as the automatic pixel-based preliminary classification first stage of a two-stage stratified hierarchical RS-IUS instantiation. Fourth, to sustain the feasibility of the new downscaled LSRC implementations, a novel vegetation spectral index is introduced and discussed. In Part II of this paper, experimental results are presented and discussed for the entire SRC family of classifiers.

In Part I of this paper, an operational fully automated Landsat-like image spectral rule-based decision-tree classifier (LSRC), suitable for mapping radiometrically calibrated seven-band Landsat-4/-5 Thematic Mapper (TM) and Landsat-7 Enhanced TM+ (ETM+) spaceborne images [eventually synthesized from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and the Moderate Resolution Imaging Spectroradiometer (MODIS) imaging sensor] into a discrete and finite set of spectral categories, has been downscaled to properly deal with spaceborne multispectral imaging sensors whose spectral resolution overlaps with, but is inferior to Landsat's, namely: 1) Satellite Pour l'Observation de la Terre (SPOT)-4/-5, Indian Remote Sensing Satellite (IRS)-1C/-1D/-P6 Linear Imaging Self-Scanner (LISS)-III, and IRS-P6 Advanced Wide Field Sensor (AWiFS); 2) National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) and Meteosat Second Generation (MSG); 3) Environmental Satellite (ENVISAT) Advanced Along-Track Scanning Radiometer (AATSR); 4) GeoEye-1, IKONOS-2, QuickBird-2, OrbView-3, TopSat, KOrean MultiPurpose SATellite (KOMPSAT)-2, FORMOsa SATellite (FORMOSAT)-2, Advanced Land Observing Satellite (ALOS) Advanced Visible and Near Infrared Radiometer type 2 (AVNIR-2), RapidEye, WorldView-2, PLEIADES-1/-2, and SPOT-6/-7; and 5) Disaster Monitoring Constellation (DMC), IRS-P6 LISS-IV, and SPOT-1/-2. LSRC, together with its five downscaled versions, identified, respectively, as the four-band SPOT-like SRC (SSRC), the four-band AVHRR-like SRC (AVSRC), the five-band AATSR-like SRC (AASRC), the four-band IKONOS-like SRC (ISRC), and the three-band DMC-like SRC (DSRC), form the so-called integrated SRC system of systems. In this paper, first, the classification accuracy and robustness to changes in the input data set of SSRC, AVSRC, AASRC, ISRC, and DSRC are assessed, both qualitatively and quantitatively, in comparison with LSRC's. Next, ongoing and future SRC applications are presented and discussed. They encompass: 1) the implementation of operational two-stage stratified hierarchical Remote Sensing (RS) image understanding systems discussed in Part I of this paper; 2) the integration of near real-time satellite mapping services with Internet map servers; and 3) the development of a new approach to semantic querying of large-scale multisensor image databases. These experimental results and application examples prove that the integrated SRC system of systems is operational, namely, it is effective, near real-time, automatic, and robust to changes in the input data set. Therefore, SRC appears eligible for use in operational satellite-based measurement systems such as those envisaged by the ongoing international Global Earth Observation System of Systems (GEOSS) Programme and the Global Monitoring for Environment and Security (GMES) system project.

In the above titled two-part paper (ibid., vol. 48, no. 3, pp. 1299-1325 and 1326-1354, Mar. 2010), there are errors which are corrected here.

The variogram image extracted from the panchromatic band of a QuickBird image has been analyzed to discriminate between similar spectral vegetated land use classes, such as forest and dense olive groves, once spectral features have been exploited. Such images seem to reproduce the direction and inter-row spacing of oriented texton-based textures present in the original input image. The variogram based texture algorithm (VBTA), recently proposed in literature for analyzing aerial images, tries to determine if texture in a region is not oriented, as in the case of forested and grassland areas or is oriented in one or two main directions, as in the case of vineyards and olive groves, respectively. If texture is oriented, the number of orientations, the orientation angle and periods (i.e. the distance between textons in each direction) are extracted, the latter depending on agricultural practices. This paper concentrates on: 1) the experimental selection of some variables useful for variogram analysis: i) the size of the input image moving window for variogram calculation; ii) some threshold values to be used for variogram normalization and for testing the presence/absence of any directionality in the variogram image; 2) The methodology for orientation and period estimation from the cumulative variance of the variogram image. The results obtained for a Southern Italy area characterized by intensive olive trees cultivation reveals the effectiveness of VBTA for QuickBird images.

This paper presents a fuzzy expert system for the creation of landslide possibility maps using change of land-use data from Earth observation, as well as historical, rainfall, and earthquake data stored in a geographic information system, as input. The difference with other systems is in the use of change (differential) input data. The method is tested with 16 documented landslides. The fuzzy neural network (NN) developed can predict the crowns of 13 out of the 16 landslides to be among the 5% most at-risk pixels that are identified in the area of study, which covers 100 km2. The fuzzy expert system considers the rules that increase the possibility of a landslide, as supplied by experts, and expresses them in the form of an empirical algebraic formula. It then fuzzifies the various thresholds they rely on and, in conjunction with uncertainties that are reported by the classifier that decides the land-use change, produces a fuzzy algebraic formula that may be used to identify the range of uncertainty in the possibility of a landslide in terms of the ranges of uncertainty in the input variables. This formula is used to train an Ishibuchi fuzzy NN, which has been designed to capture uncertainty in the rules and uncertainty in the input variables. It is this Ishibuchi NN that acts as a fuzzy expert system.

Although underestimated in practice, the small/unrepresentative sample problem is likely to affect a large segment of real-world remotely sensed (RS) image mapping applications where ground truth knowledge is typically expensive, tedious, or difficult to gather. Starting from this realistic assumption, subjective (weak) but ample evidence of the relative effectiveness of existing unsupervised and supervised data labeling systems is collected in two RS image classification problems. To provide a fair assessment of competing techniques, first the two selected image datasets feature different degrees of image fragmentation and range from poorly to ill-posed. Second, different initialization strategies are tested to pass on to the mapping system at hand the maximally informative representation of prior (ground truth) knowledge. For estimating and comparing the competing systems in terms of learning ability, generalization capability, and computational efficiency when little prior knowledge is available, the recently published data-driven map quality assessment (DAMA) strategy, which is capable of capturing genuine, but small, image details in multiple reference cluster maps, is adopted in combination with a traditional resubstitution method. Collected quantitative results yield conclusions about the potential utility of the alternative techniques that appear to be realistic and useful in practice, in line with theoretical expectations and the qualitative assessment of mapping results by expert photointerpreters.

Based on purely spectral-domain prior knowledge taken from the remote sensing (RS) literature, an original spectral (fuzzy) rule-based per-pixel classifier is proposed. Requiring no training and supervision to run, the proposed spectral rule-based system is suitable for the preliminary classification (primal sketch, in the Marr sense) of Landsat-5 Thematic Mapper and Landsat-7 Enhanced Thematic Mapper Plus images calibrated into planetary reflectance (albedo) and at-satellite temperature. The classification system consists of a modular hierarchical top-down processing structure, which is adaptive to image statistics, computationally efficient, and easy to modify, augment, or scale to other sensors' spectral properties, like those of the Advanced Spaceborne Thermal Emission and Reflection Radiometer and of the Satellite Pour l'Observation de la Terre (SPOT-4 and -5). As output, the proposed system detects a set of meaningful and reliable fuzzy spectral layers (strata) consistent (in terms of one-to-one or many-to-one relationships) with land cover classes found in levels I and II of the U.S. Geological Survey classification scheme. Although kernel spectral categories (e.g., strong vegetation) are detected without requiring any reference sample, their symbolic meaning is intermediate between those (low) of clusters and segments and those (high) of land cover classes (e.g., forest). This means that the application domain of the kernel spectral strata is by no means alternative to RS data clustering, image segmentation, and land cover classification. Rather, prior knowledge-based kernel spectral categories are naturally suitable for driving stratified application-specific classification, clustering, or segmentation of RS imagery that could involve training and supervision. The efficacy and robustness of the proposed rule-based system are tested in two operational RS image classification problems

This paper deals with the problem of badly posed image classification. Although underestimated in practice, bad-posedness is likely to affect many real-world image classification tasks, where reference samples are difficult to collect (e.g., in remote sensing (RS) image mapping) and/or spatial autocorrelation is relevant. In an image classification context affected by a lack of reference samples, an original inductive learning multiscale image classifier, termed multiscale semisupervised expectation maximization (MSEM), is proposed. The rationale behind MSEM is to combine useful complementary properties of two alternative data mapping procedures recently published outside of image processing literature, namely, the multiscale modified Pappas adaptive clustering (MPAC) algorithm and the sample-based semisupervised expectation maximization (SEM) classifier. To demonstrate its potential utility, MSEM is compared against nonstandard classifiers, such as MPAC, SEM and the single-scale contextual SEM (CSEM) classifier, besides against well-known standard classifiers in two RS image classification problems featuring few reference samples and modestly useful texture information. These experiments yield weak (subjective) but numerous quantitative map quality indexes that are consistent with both theoretical considerations and qualitative evaluations by expert photointerpreters. According to these quantitative results, MSEM is competitive in terms of overall image mapping performance at the cost of a computational overhead three to six times superior to that of its most interesting rival, SEM. More in general, our experiments confirm that, even if they rely on heavy class-conditional normal distribution assumptions that may not be true in many real-world problems (e.g., in highly textured images), semisupervised classifiers based on the iterative expectation maximization Gaussian mixture model solution can be very powerful in practice when: 1) there is a lack of reference samples with respect to the problem/model complexity and 2) texture information is considered negligible (i.e., a piecewise constant image model holds).

This work focuses on two challenging types of problems related to quality assessment and comparison of thematic maps generated from remote sensing (RS) images when little or no ground truth knowledge is available. These problems occur when: (1) competing thematic maps, generated from the same input RS image, assumed to be available, must be compared, but no ground truth knowledge is found to assess the accuracy of the mapping problem at hand, and (2) the generalization capability of competing classifiers must be estimated and compared when the small/unrepresentative ground truth problem affects the RS inductive learning application at hand. Specifically focused on badly posed image classification tasks, this paper presents an original data-driven (i.e., unsupervised) thematic map quality assessment (DAMA) strategy complementary (not alternative) in nature to traditional supervised map accuracy assessment techniques, driven by the expensive and error-prone digitization of ground truth knowledge. To compensate for the lack of supervised regions of interest, DAMA generates so-called multiple reference cluster maps from several blocks of the input RS image that are clustered separately. Due to the unsupervised (i.e., subjective) nature (ill-posedness) of data clustering, DAMA provides no (absolute) map accuracy measure. Rather, DAMA's map quality indexes are to be considered unsupervised (i.e., subjective) relative estimates of labeling and segmentation consistency between every competing map at hand and the set of multiple reference cluster maps. In two badly posed RS image mapping experiments, DAMA's map quality measures are proven to be: (1) useful in the relative comparison of competing mapping systems; (2) consistent with theoretical expectations; and (3) in line with mapping quality criteria adopted by expert photointerpreters. Documented limitations of DAMA are that it is intrinsically heuristic due to the subjective nature of the clustering problem, and like any evaluation measure, it cannot be injective.

A Bayesian approach to detect land cover changes is presented. Comparison of classification maps performed at different dates is a straightforward tool for obtaining changes occurred in between. Unfortunately the direct comparison of classification maps is characterized by an high probability of occurrence of false positive change events. To overcome this problem, it is necessary that the process might be driven by the a priori knowledge about considered scene. In this framework, the change detection is seen as a data fusion task where the a priori class transitions drives a multitemporal classification. The validity of the proposed method is confirmed by the results, which are discussed in this paper. In fact, when compared with an independent classification scheme, the approach based on a priori knowledge permits to achieve a higher overall accuracy without loosing in efficiency.

In this work, unsupervised change detection techniques, based on three different way to compare images, are presented. Two Landsat TM registered and corrected multi-spectral images, acquired on the same geographical area on 18 May 1996 and 21 May 1997, have been used. In the first comparison technique, for each pair of corresponding pixels, the spectral change vector has been computed as the squared difference in the features vectors at the two times. In the second method, the difference image has been computed using, pixel by pixel, a chi square transformation. The third technique is based on the application of a Self-Organizing Map (SOM) neural network to clusterize the two images before comparison. The three obtained difference images has been then analyzed by using a fully automatic thresholding method exploiting the expectation-maximization (EM) algorithm. The experimental results obtained for the three difference images are comparable, showing a reliable robustness of the unsupervised approach, and only few change are detected on the analyzed scene. Moreover, the experimental results have been compared with a change detection map computed by using a supervised technique, obtaining a good agreement between unsupervised and supervised results that confirms the reliability of the considered approach. The encouraging obtained results allow to use the so-computed percentage value of changes as probability of class transitions in input to a Bayesian supervised change detection method, as presented in a companion paper by the same authors. In this framework, the unsupervised approach may be used to support supervised techniques, providing land cover transitions that can be used as guess values

Slope instability studies seem to recognize and group a number of potential superficial slide-producing agents which might be directly detected and monitored from Earth Observation (EO) data. The attention of this work is focused on man's activity induced surface changes, such as deforestation, urban expansion, artificial structures construction. An historical set of fourteen multi-temporal optical Landsat TM images have been considered. The main objective of the work is to verify the advantages and limitations of conventional space-borne RS data to provide change maps on areas including unstable slopes. A supervised change detection technique is preferred to an unsupervised technique since the former can provide a change image containing useful information not only on the place were a transition occurred, but also on the specific classes involved in the transitions between two dates. The change image is used to extract class-conditional transition probabilities and evaluate class-specific trends of change. Four classes and their transitions have been considered in the analysis: (1) arboreous land, (2) agricultural land, (3) barren land, (4) artificial structures. The percentage values of the total number of changed pixels for each map pair is also correlated with known landslides events occurred in the considered period. A correlation value of 0.8 is obtained. This paper discusses the results obtained on a test site located in Regione Abruzzo, Southern Italy, affected by slope instability phenomena

Photo interpretation by human experts has been the major data source for urban planning and monitoring applications. However, images collected from a space based sensor which combines reasonably good spectral and spatial resolution, could provide an useful tool for automatic monitoring of urban area changes. With this aim, in this paper, a data fusion technique, based on a RGB - HIS transformation, has been adopted for combining high spatial resolution panchromatic satellite with multi-spectral low resolution IKONOS II images. Moreover , textural information, which characterizes urban area, has been extracted with the use of a filter for the edge extraction. An MLP classifier has been trained to produce a labelled image with great accuracy in test even if a limited training set has been used. For a photo interpreter, the results reveal a good feasibility of the classified image for monitoring the presence of changes in urban areas, useful for a cartographic updating. different years. The percentage of correctly classified objects was 89%, whereas the percentage of correct object changes was equal to 83 %, with a false alarm rate of 5%. In (3) a parallelepiped supervised classification algorithm is used to obtain a land cover map characterized by seven classes from two pan-sharpened multi-spectral images at 1m resolution. Overall accuracy values of 75-83% are obtained on test data by using a PS-MS / RGB band composition and an RGB / NIR band composition, respectively. In (4) a neuro-fuzzy classifier based on a set of IF- Then-rules is compared with a Back Propagation neural network and a Maximum Likelihood (ML) classifier to produce a land cover map from IKONOS data on the Korean peninsula, characterized by mixed composition areas. The neuro-fuzzy classifier was more accurate than other classifiers on mixed composition areas, whereas the maximum likelihood performed better on areas such as roads. As input features, the authors considered the four multi-spectral band at low spatial resolution. The usefulness of IKONOS imagery for classification of urban and suburban scenes was also investigated in (5), where a hierarchical fuzzy classification techniques is proposed to improve the classification accuracy obtained by a ML traditional classifier when using only the MS bands of IKONOS data. The ML performance of about 81% was in fact increased to 88%, by using as input to the hierarchical classifier some textural features, extracted from the PAN band, beside the ML classified image. The objective of this work was twofold. First, to validate the feasibility of the RGB-HSI data fusion approach to assimilate the information derived from IKONOS high spatial / low spectral resolution data and the spectral information from low spatial / high spectral resolution. Second, to fully exploit the contextual information of IKONOS PAN data for the automatic classification of urban settlements. Two areas in Southern Italy, characterized by a different typical landscape, were selected for the study. The first is a country area with little villages, tourist facilities and isolated holiday houses

This paper presents a comparative evaluation between a classification strategy based on the combination of the outputs of a neural (NN) ensemble and the application of Support Vector Machine (SVM) classifiers in the analysis of remotely sensed data. Two sets of experiments have been carried out on a benchmark data set. The first set concerns the application of linear and non linear techniques to the combination of the outputs of a Multilayer Perceptron (MLP) neural network ensemble. In particular, the Bayesian and the error correlation matrix approaches are used for coefficient selection in the linear combination of the network's outputs. A MLP module is used for the non linear outputs combination. The results of linear and non linear combination schemes are compared and discussed versus the performance of SVM classifiers. The comparative analysis evidences that the nonlinear, MLP based, combination provides the best results among the different combination schemes. On the other hand, better performance can be obtained with SVM classifiers. However, the complexity of the SVM training procedure can be considered a limitation for SVMs application to real-world problems.

In this work, the results obtained in the classification of a multi-source - multi-temporal remote sensed data set by means of a distributed neuro-fuzzy system are compared with the results of a traditional centralized neural classification system, based on a single multilayer perceptron (MLP) neural network module. The distributed system is composed by a set of neural classifiers, whose partial results were combined with both Sugeno and Choquet fuzzy integrals. Two classification experiments were carried out with the distributed system. In the first experiment, each neural module of the distributed system used the same learning rule but was trained with a subset of the input features, i.e., a specific spectral band. In the second experiment, the neural modules of the system were trained with the same complete set of input features available for each training pixel, but consisted of MLP networks characterized by different specific topologies or different neural algorithms. The results show that larger improvements can be obtained by combining more independent classifiers. The Choquet fuzzy integral provided better performance than Sugeno fuzzy integral. The centralized system, based on a single MLP module, provided the best classification performance.

Mixed pixels are a major source of inconvenience in the classification of remotely sensed data. This paper compares MLP with so-called neuro-fuzzy algorithms in the estimation of pixel component cover classes. Two neuro-fuzzy networks are selected from the literature as representatives of soft classifiers featuring different combinations of fuzzy set-theoretic principles with neural network learning mechanisms. These networks are: 1) the fuzzy multilayer perceptron (FMLP) and 2) a two-stage hybrid (TSH) learning neural network whose unsupervised first stage consists of the fully self-organizing simplified adaptive resonance theory (FOSART) clustering model, FMLP, TSH, and MLP are compared on CLASSITEST, a standard set of synthetic images where per-pixel proportions of cover class mixtures are known a priori. Results are assessed by means of evaluation tools specifically developed for the comparison of soft classifiers. Experimental results show that classification accuracies of FMLP and TSH are comparable, whereas TSH is faster to train than FMLP. On the other hand, FMLP and TSW outperform MLP when little prior knowledge is available for training the network, i.e., when no fuzzy training sites, describing intermediate label assignments, are available.

Clustering algorithms aim at modeling fuzzy (i.e., ambiguous) unlabeled patterns efficiently. Our goal is to propose a theoretical framework where the expressive power of clustering systems can be compared on the basis of a meaningful set of common functional features. Part I of this paper reviews the following issues related to clustering approaches found in the literature: relative (probabilistic) and absolute (possibilistic) fuzzy membership functions and their relationships to the Bayes rule, batch and on-line learning, prototype editing schemes, growing and pruning networks, modular network architectures, topologically perfect mapping, ecological nets and neuro-fuzziness. From this discussion an equivalence between the concepts of fuzzy clustering and soft competitive learning in clustering algorithms is proposed as a unifying framework in the comparison of clustering systems. Moreover, a set of functional attributes is selected for use as dictionary entries in the comparison of clustering algorithms, which is the subject of part II of this paper.

For pt.I see ibid., p.775-85. In part I an equivalence between the concepts of fuzzy clustering and soft competitive learning in clustering algorithms is proposed on the basis of the existing literature. Moreover, a set of functional attributes is selected for use as dictionary entries in the comparison of clustering algorithms. In this paper, five clustering algorithms taken from the literature are reviewed, assessed and compared on the basis of the selected properties of interest. These clustering models are (1) self-organizing map (SOM); (2) fuzzy learning vector quantization (FLVQ); (3) fuzzy adaptive resonance theory (fuzzy ART); (4) growing neural gas (GNG); (5) fully self-organizing simplified adaptive resonance theory (FOSART). Although our theoretical comparison is fairly simple, it yields observations that may appear parodoxical. First, only FLVQ, fuzzy ART, and FOSART exploit concepts derived from fuzzy set theory (e.g., relative and/or absolute fuzzy membership functions). Secondly, only SOM, FLVQ, GNG, and FOSART employ soft competitive learning mechanisms, which are affected by asymptotic misbehaviors in the case of FLVQ, i.e., only SOM, GNG, and FOSART are considered effective fuzzy clustering algorithms.

Fuzzy learning vector quantization (FLVQ), also known as the fuzzy Kohonen clustering network, was developed to improve performance and usability of on-line hard-competitive Kohnen's vector quantization and soft-competitive self organizing map (SOM) algorithms. The FLVQ effectiveness seems to depend on the range of change of the weighting exponent m(t). In the first part of this work, extreme m(t) values (1 and 1, respectively) are employed to investigate FLVQ asymptotic behaviors. This analysis shows that when m(t) tends to either one of its extremes, FLVQ is affected by trivial vector quantization, which causes centroids collapse in the grand mean of the input data set. No analytical criterion has been found to improve the heuristic choice of the range of m(t) change. In the second part of this paper, two FLVQ and SOM classification experiments of remote sensed data are presented. In these experiments the two nets are connected in cascade to a supervised second stage, based on the delta rule. Experimental results confirm that FLVQ performance can be greatly affected by the user's definition of the range of change of the weighting exponent. Moreover, FLVQ shows instability when its traditional termination criterion is applied. Empirical recommendations are proposed for the enhancement of FLVQ robustness. Both the analytical and the experimental data reported seem to indicate that the choice of the range of m(t) change is still open to discussion and that alternative clustering neural-network approaches should be developed to pursue during training: 1) monotone reduction of the neurons' learning rate and 2) monotone reduction of the overlap among neuron receptive fields.

Prospective authors are requested to submit new, unpublished manuscripts for inclusion in the upcoming event described in this call for papers.