Search

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 Pins 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. Index Terms--Data clustering, fuzzy rule, fuzzy set (FS), generalization capability, image classification, image color analysis, image segmentation, one-class classifier, prior knowledge, remotely sensed imagery, spectral information, supervised and unsupervised learning from finite data.

The original heuristic multiscale semisupervised expectation maximization (MSEM) is proposed as an improvement over existing nonstandard data mapping systems like SEM, contextual SEM and modified Pappas adaptive clustering, in the badly posed classification of piecewise constant or slowly varying color images. The results have revealed that although they rely on heavy class-conditional normal distribution assumptions that are not true in many real-world problems, semisupervised classifiers using EM can be very powerful in practice.

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. Index Terms--Badly posed classification, competing classifier evaluation, clustering, curse of dimensionality, generalization capability, image labeling, inductive learning, map accuracy assessment, remotely sensed (RS) imagery, semilabeled sampies, semisupervised learning, supervised learning, unsupervised learning.

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. Index Terms--Badly posed classification, clustering, competing classifier evaluation, generalization capability, image mapping, quality assessment of maps, remotely sensed images, resampling techniques for estimating statistics, sampling techniques for reference data selection, supervised learning, unsupervised learning.

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 [infinity], 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. Index Terms - Fuzzy clustering, growing neural-network, hybrid classification system, probabilistic and possibilistic membership, self-organizing map, soft and hard competitive learning, topology preserving mapping.

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 TSH 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. Index Terms--Classification accuracy measure, clustering, neuro-fuzzy classifier, per-pixel spectral unmixing, soft and hard classification, supervised and unsupervised learning.

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 member. ship 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 [1]. Index Terms--Ecological net, fuzzy clustering, modular architecture, relative and absolute membership function, soft and hard competitive learning, topologically correct mapping.