Skip to main content

Über dieses Buch

Since 1994 the European Commission has been supporting activities under the Environment and Climate programme of research and technological de­ velopment, with the aim of developing cost-effective applications of satellite Earth observation (EO) for both environmental monitoring and research. This action has included support to methodological research, aimed at the development and evaluation of new techniques forming part ofthe chain of processing needed to transform data into useful information. Wherever appropriate, the Commission has emphasised the coordination of ongoing research funded at the national level, through the mechanism of concerted actions. Concerted actions are flexible and efficient means to marshal efforts at the European level for a certain period. They are proposed by groups of researchers active in a given field who have identified the added value to be gained by European cooperation, whilst continuing to pursue their own individual projects. In view of the rapid developments in the field of neural network over the last 10 years, together with the growing interest ofthe Earth observation community in this approach as a tool for data interpretation, the Commission decided in 1995 to support the concerted action COMPARES, following a proposal from a group of acknowledged European experts.




Neural network or connectionist algorithms have made an enormous impact in the field of signal processing over the last decade. Although few would regard them as the perfect answer to all pattern recognition problems, there is little doubt that they have contributed significantly to the solution of some of the most difficult ones. Trainable networks of primitive processing elements have been shown to be capable of describing and modelling systems of great complexity without the necessity of building parameterised statistical descriptions. Such capability has led to considerable and constantly growing interest from the remote sensing community. From early beginnings in the late 1980’s, neural network algorithms are now being explored for a wide range of uses in Earth observation. In many cases these uses are still experimental or at the pre-operational stage.
Ioannis Kanellopoulos, Graeme G. Wilkinson, Fabio Roli, James Austin

Open Questions in Neurocomputing for Earth Observation

Neural network usage in remote sensing has grown dramatically in recent years - mostly for classification. However neural systems are susceptible to problems such as unpredictability, over-fitting and chaos which render them unsatisfying for many potential users. In the classification context they have yet to show any major improvement over conventional statistical algorithms in generalisation to large geographical areas (regional - continental scale). Apart from classification, neural computation can solve a wider range of problems in remote sensing than have been explored so far -especially related to geometrical processing and signal inversion. Further important issues are raised which should form the basis for future research such as the need for very large networks (which are sensitive to geographical and temporal context), the requirement for special purpose hardware, and the need for better user-adapted systems.
Graeme G. Wilkinson

A Comparison of the Characterisation of Agricultural Land Using Singular Value Decomposition and Neural Networks

Methods are defined and tested for the characterisation of agricultural land from multi-spectral imagery, based on Singular Value Decomposition (SVD) and Artificial neural networks (ANN). The SVD technique, which bears a close resemblance to multivariate statistic techniques, has previously been successfully applied to problems of signal extraction for marine data [1] and forestry species classification [2].
In this study the two techniques are used as a classifier for agricultural regions, using airborne Daedalus ATM data, with lm resolution. The specific region chosen is an experimental research farm in Bavaria, Germany. This farm has a large number of crops, within a very small region and hence is not amenable to existing techniques. There are a number of other significant factors which render existing techniques such as the maximum likelihood algorithm less suitable for this area. These include a very dynamic terrain and tessellated pattern soil differences, which together cause large variations in the growth characteristics of the crops.
Both the SVD and ANN techniques are applied to this data set using a multistage classification approach. Typical classification accuracy’s for the techniques are of the order of 85–100%. Preliminary results indicate that the methods provide fast and efficient classifiers with the ability to differentiate between crop types such as Wheat, Rye, Potatoes and Clover.
S. Danaher, G. Herries, T. Selige, M. Mac Súirtán

Land Cover Mapping from Remotely Sensed Data with a Neural Network: Accommodating Fuzziness

Neural networks are attractive for the supervised classification of remotely sensed data. There are, however, many problems with their use, restricting the realisation of their full potential. This article focuses on the accommodation of fuzziness in the classification procedure. This is required if the classes to be mapped are continuous or if there is a large proportion of mixed pixels. A continuum of fuzzy classifications was proposed and it is shown that a neural network may be configured at any point along this continuum, from a completely-hard to a fully-fuzzy classification. Examples of fuzzy classifications are given illustrating the potential for mapping continuous classes and reducing the mixed pixel problem.
Giles M. Foody

Geological Mapping Using Multi-Sensor Data: A Comparison of Methods

Landsat TM and SIR-C SAR data are used in a comparative test of the performance of a statistical classifier, the maximum likelihood (ML) procedure, and two neural networks, a multi-layer feed-forward network (F-NN) and a Self-Organising Map (SOM). Using spectral features alone, performance (as measured by comparison with unpublished field maps) of all three methods is poor, with overall accuracies of less than 60%. The F-NN performs best. When texture measures are added, overall classification accuracy is improved, with the Grey Level Cooccurrence Matrix (GLCM) approach showing the best result (overall accuracy of almost 70%).
The ML procedure requires less than two minutes’ computing time for the 10242 test image, whereas the unsupervised learning stage of the SOM required of the order of 70 hours. Design and evaluation of the F-NN was also time-consuming. The cost of using the GCLM procedure to derive texture features is proportional to the number of grey levels in the image (256 in the case of Landsat TM and SIR-C SAR). Reducing the number of grey levels to 64 by means of an equalising algorithm proved to have little effect upon performance.
Paul M. Mather, Brandt Tso, Magaly Koch

Application of Neural Networks and Order Statistics Filters to Speckle Noise Reduction in Remote Sensing Imaging

A novel approach to suppression of speckle noise in remote sensing imaging based on a combination of segmentation and optimum L-filtering is presented. With the aid of a suitable modification of the Learning Vector Quantizer (LVQ) neural network, the image is segmented in regions of (approximately) homogeneous statistics. For each of the regions a minimum mean-squared-error (MMSE) L-filter is designed, by using the histogram of grey levels as an estimate of the parent distribution of the noisy observations and a suitable estimate of the (assumed constant) original signal in the corresponding region. Thus, a bank of L-filters results, with each of them corresponding to and operating on a different image region. Simulation results are presented, which verify the (qualitative and quantitative) superiority of our technique over a number of commonly used speckle filters.
E. Kofidis, S. Theodoridis, C. Kotropoulos, I. Pitas

Neural Nets and Multichannel Image Processing Applications

The application of neural network technology to multichannel image processing is presented in this paper. Topics such as image restoration, segmentation, transformation and compression are discussed, covering a wide range of image processing and analysis areas. The problems are converted to optimisation or interpolation problems through the appropriate mathematical interface. This alternative interpretation enables the application of well-known neural network structures, perfectly suited for such purposes due to the accuracy and high speed of computation. The proposed framework handles multichannel data in a compact form and, thus, it is directly applicable to remote sensing.
Vassilis Gaganis, Michael Zervakis, Manolis Christodoulou

Neural Networks for Classification of Ice Type Concentration from ERS-1 SAR Images

Classical Methods versus Neural Networks
This paper describes a minor part of the work done in connection with a preliminary investigation of a neural network’s capability to classify ice types. It includes a short review of earlier used techniques, implementation of different neural networks and results from various experiments with these networks. The estimation of ice type concentrations from Synthetic Aperture Radar (SAR) images has been investigated for several years, see e.g. [9]. The classification estimation has been performed by training a Bayesian Maximum Likelihood Classifier (BMLC) [8] with a classification rate about 80%. The neural networks considered are all of the feed-forward type. For training, different learning algorithms and error functions are used. Both pruning and construction algorithms are used to get an optimal architecture. Experiments showed that almost any kind of neural network, using a Standard Back-Propagation (Std_BP) learning algorithm for minimising the Mean Square Error (MSE), is able to perform better than the BMLC. The reason is that the neural network is able to use a larger training set than the BMLC.
Jan Depenau

A Neural Network Approach to Spectral Mixture Analysis

Imaging spectrometers acquire images in many narrow spectral bands. Because of the limited spatial resolution, often more than one ground cover category is present in a single pixel. In spectral mixture analysis the fractions of the ground cover categories present in a pixel are determined, assuming a linear mixture model. In this paper neural network methods which are able to perform this analysis are considered. Methods for the construction of training and test data sets for the neural network are given. Using data from 3 spectrometers with 6, 30 and 220 bands and 3 or 4 ground cover categories, it is shown that a back-propagation neural network with one hidden layer is able to learn the relation between the intensities of a pixel and its ground cover fractions. The distributions of the differences between true and calculated fractions show that a neural network performs the same or better than a conventional least squares with covariance matrix method. The calculation of the fractions by a neural network is much faster than by the least squares methods, training of the neural networks requires however a large amount of computer time.
Theo E. Schouten, Maurice S. klein Gebbinck

Comparison Between Systems of Image Interpretation

In this paper, we present classification methods, which have been adapted to the pattern recognition problem in the domain of aerial imagery. Three methods based on different mechanisms have been developed: rule-based system, neural networks and fuzzy classification. Our contributions are the adaptation of these methods to the concrete case of aerial images, and the quantitative comparison between the three types of methods.
Jean-Pierre Cocquerez, Sylvie Philipp, Philippe Gaussier

Feature Extraction for Neural Network Classifiers

Classification of multi-source remote sensing and geographic data by neural networks is discussed with respect to feature extraction. Several feature extraction methods are reviewed, including principal component analysis, discriminant analysis, and the recently proposed decision boundary feature extraction method. The feature extraction methods are then applied in experiments in conjunction with classification by multilayer neural networks. The decision boundary feature extraction method shows excellent performance in the experiments.
Jon A. Benediktsson, Johannes R. Sveinsson

Spectral Pattern Recognition by a Two-Layer Perceptron: Effects of Training Set Size

Pattern recognition in urban areas is one of the most challenging issues in classifying satellite remote sensing data. Parametric pixel-by-pixel classification algorithms tend to perform poorly in this context. This is because urban areas comprise a complex spatial assemblage of disparate land cover types - including built structures, numerous vegetation types, bare soil and water bodies. Thus, there is a need for more powerful spectral pattern recognition techniques, utilising pixel-by-pixel spectral information as the basis for automated urban land cover detection. This paper adopts the multi-layer perceptron classifier suggested and implemented in [5]. The objective of this study is to analyse the performance and stability of this classifier - trained and tested for supervised classification (8 a priori given land use classes) of a Landsat-5 TM image (270 × 360 pixels) from the city of Vienna and its northern surroundings - along with varying the training data set in the singletraining-site case. The performance is measured in terms of total classification, map user’s and map producer’s accuracies. In addition, the stability with initial parameter conditions, classification error matrices, and error curves are analysed in some detail.
Petra Staufer, Manfred M. Fischer

Comparison and Combination of Statistical and Neural Network Algorithms for Remote-Sensing Image Classification

In recent years, the remote-sensing community has became very interested in applying neural networks to image classification and in comparing neural networks performances with the ones of classical statistical methods. These experimental comparisons pointed out that no single classification algorithm can be regarded as a “panacea”. The superiority of one algorithm over the other strongly depends on the selected data set and on the efforts devoted to the “designing phases” of algorithms. In this paper, we propose the use of “ensembles” of neural and statistical classification algorithms as an alternative approach based on the exploitation of the complementary characteristics of different classifiers. Classification results provided by image classifiers contained in these ensembles are “merged” according to statistical combination methods. Experimental results on a multi-sensor remote-sensing data set point out that the use of classifiers ensembles can constitute a valid alternative to the development of new classification algorithms “more complex” than the present ones. In particular, we show that the combination of results provided by statistical and neural algorithms provides classification accuracies better than the ones obtained by single classifiers after long “designing phases”.
Fabio Roli, Giorgio Giacinto, Gianni Vernazza

Integrating the Alisa Classifier with Knowledge-Based Methods for Cadastral-Map Interpretation

Alisa is a learning, statistical, texture classifier for single-and multi-class classification. Its process is based on the examination of images using a set of universal (i.e., independent of the domain of the images under examination) features. Given a set of pre-classified examples, it computes a subset of these features for a small window (i.e., the analysis token) centred at each image pixel, and creates a histogram of occurrences of the distinct feature values in the training data. After training is completed, given an unknown image Alisa examines this in the same way, and generates an isomorphic image, each pixel of which represents the normality of the corresponding pixel in the input image (or, in multi-class classification, the class in which it belongs). In this paper, we discuss the integration of Alisa with knowledge-based methods for recognising line thickness in cadastral maps.
Eleni Stroulia, Rudolf Kober

A Hybrid Method for Preprocessing and Classification of SPOT Images

In this paper, we present a hybrid method for preprocessing and classification of satellite images. The preprocessing consists of computing texture measures of the images and initialising the probabilities of pixels belonging to different land-cover classes. The objective of the preprocessing is twofold: increasing discrimination power and removing irrelevant characteristics. The classification process consists of assigning a class to each pixel, with a special interest in detecting urban areas as completely as possible with the aid of a priori knowledge. This interest stems from the possible requirement of detecting urban areas on satellite images (even small villages in the countryside) while ignoring some classes (such as parks) in cities. We shall show how this requirement is translated into constraints imposed in our classification process. Experimental results are illustrated through a SPOT image containing a coastal town.
Shan Yu, Konrad Weigl

Testing some Connectionist Approaches for Thematic Mapping of Rural Areas

An overview of several supervised classification methodologies applied at the research centre CVRM to automatic classification of remote sensing data is presented. Some classification methodologies based on Artificial Neural Networks are roughly described, with emphasis on Multilayer Feed-forward Networks (MLFN) trained with the Back-propagation algorithm and on Probabilistic Neural Networks (PNN). A technique under study at the CVRM for reducing training time, conjugating Principal Components Analysis and Genetic Algorithms, is outlined. Finally a case study is presented, describing an application of a MLFN and a PNN to a pixel based multispectral (6 bands) Landsat TM data from the rural Moura-Ficalho area (South of Portugal). The results were mapped and the performances compared, by means of confusion matrices and also by the computing time used, with a typical parametric approach such as the classical Maximum Likelihood Classifier and with field work interpretation.
Leopoldo Cortez, Fernando Durão, Vitorino Ramos

Using Artificial Recurrent Neural Nets to Identify Spectral and Spatial Patterns for Satellite Imagery Classification of Urban Areas

The majority of techniques used for satellite imagery classification usually perform poorly on discriminating urban land use classes, either because they have similar spectral signatures or because the patterns they exhibit are broader than satellite image pixels. In this paper we tackle a new classification methodology, based on spectral and spatial pattern analysis using artificial neural networks. First a self-organising classifier splits the spectrum of individual pixels on spectrally pure land cover classes. Then a second classifier self-organises critical regions of adjustable topology on the resultant image, and automatically classifies it into land use classes. Both classifiers are implemented on artificial recurrent neural networks inspired by Carpenter and Grossberg’s adaptive resonance theory (ART).
Sara Silva, Mario Caetano

Dynamic Segmentation of Satellite Images Using Pulsed Coupled Neural Networks

Pulsed Coupled Oscillatory Neural Networks are examined for application to image analysis. Adapting to biological constraints, a pulsed coupled network using an Integrate and Fire model with dynamical synapses is examined to perform image segmentation based on synchronisation on the firing time-of neurons which are in the same region. To enhance synchronisation behaviour an avalanche-type dynamic is introduced. This dynamic allows us to perform segmentation by synchronisation in the transient regime, without having to reach a stable stationary regime which is expensive in terms of computation time. Using the pulsed time-of-arrival as the information carrier, the image is reduced to a time signal which allows an intelligent filtering using feedback. A multi-layer implementation of the model is presented that shows good segmentation results and is easily adaptable to multispectral imagery.
X. Clastres, M. Samuelides, G. L. Tarr

Non-Linear Diffusion as a Neuron-Like Paradigm for Low-Level Vision

Remote sensing puts high demands on image processing. It calls for state-of-the-art algorithms, e.g. neural networks. However, neural nets usually work on preprocessed data and the preprocessing steps themselves have proved difficult to implement with NNs. Here a NN-like paradigm for low-level image processing is presented, that is based on the evolution of coupled, non-linear diffusion equations. The illustrations are focussed on feature preserving noise reduction, but the framework is more general.
M. Proesmans, L. J. Van Gool, P. Vanroose

Application of the Constructive Mikado-Algorithm on Remotely Sensed Data

Finding an optimal architecture for a neural network, i.e. an optimal number and size of layers, is an open problem. With the Mikado-algorithm we present a new method to construct the network architecture in the course of learning. With two examples, classification and structure detection from remotely sensed data, we demonstrate the capabilities of the Mikado-algorithm. This algorithm provides good generalisation in the presence of mixed pixels, delivering small networks for high-dimensional problems and a new way of interpreting the network generalisation ability.
C. Cruse, S. Leppelmann, A. Burwick, M. Bode

A Simple Neural Network Contextual Classifier

In this paper we describe a neural network used to make a simple contextual classifier using a two layer feed-forward network. The best number of hidden units is chosen by training a network with too many hidden units. We then prune the network using Optimal Brain Damage (OBD). The pruned networks have a better generalisation error because they only have the weights that reflect the structure of the data and not the noise. We study the possibility of using a Network Information Criterion (NIC) to decide when to stop pruning. When we use NIC we can estimate the test error of a network without using an independent validation set.
As a case study we use a four band Landsat-2 Multispectral Scanner (MSS) image from southern Greenland. To classify a pixel in the non-contextual case we use the four variables from the MSS bands only. In the simple contextual case we augment the feature vector with the four mean values of the MSS bands from the four nearest neighbours. We notice an increase in the number of correct classified pixels when using the contextual classifier. Also, the application of the simple contextual classifier gives a small overall increase in the posterior probability.
Jens Tidemann, Allan Aasbjerg Nielsen

Optimising Neural Networks for Land Use Classification

In this paper we present a fully automatic and computationally efficient algorithm for optimising multilayer perceptron classifiers. The approach involves two procedures: adaptation (training) and selection. The first procedure adaptively changes the weights of the network. The selection procedure performs the elimination of some of the hidden units (weights). By iteratively combining these two procedures we achieve a controlled way of training and modifying neural networks, which balances accuracy, learning time, and complexity of the resulting network. We demonstrate our method on the problem of multispectral Landsat image classification. We compare our results with a hand designed multi-layer perceptron and a Gaussian maximum likelihood classifier on the same data. Our method produces a better classification accuracy with a smaller number of hidden units than the hand designed network.
Horst Bischof, Aleš Leonardis

High Speed Image Segmentation Using a Binary Neural Network

In the very near future large amounts of Remotely Sensed data will become available on a daily basis. Unfortunately, it is not clear if the processing methods are available to deal with this data in a timely fashion. This paper describes research towards an approach which will allow a user to perform a rapid pre-search of large amounts of image data for regions of interest based on texture. The method is based on a novel neural network architecture (ADAM) that is designed primarily for speed of operation by making use of computationally simple pre-processing and only uses Boolean operations in the weights of the network. To facilitate interactive use of the network, it is capable of rapid training. The paper outlines the neural network, its application to RS data in comparison with other methods, and briefly describes a fast hardware implementation of the network.
Jim Austin

Efficient Processing and Analysis of Images Using Neural Networks

A neural network based approach for efficient processing and analysis of images in remote sensing applications is proposed in this paper. Classification, segmentation, coding and interpretation of e.g., Landsat image data are tasks which can take advantage of this approach. First, we use multi-resolution analysis of the images in order to obtain image representations of lower sizes, to which neural networks can be applied more effectively. It is shown that auto-associative neural networks can be used to perform the multi-resolution analysis in an optimal way. Hierarchical neural networks are designed which are able to implement the analysis and classification task at different resolutions. Neural networks are also proposed as an efficient means for selecting regions of interest in the images, further reducing the image representation size which is necessary for effective analysis and classification of the images. Application of the proposed approach to specific real life remote sensing image data is currently under investigation.
Stefanos Kollias

Selection of the Number of Clusters in Remote Sensing Images by Means of Neural Networks

When processing multidimensional remote sensing data, one of the main problems is the choice for the appropriate number of clusters: despite a great amount of good algorithms for clustering, each of them works properly only when the appropriate number of clusters is selected. As an adaptive version of K-means, the Competitive Learning algorithm (CL) also has a similar crucial problem: different modifications to CL were made with the introduction of frequency sensitive competitive learning (FSCL) and rival penalised competitive learning (RPCL) recently. This last approach introduces an interesting competition mechanism but fails in the presence of real data with multiple clusters of different dimension. We present an improvement of the RPCL algorithm well adapted to work with every kind of real clustering data problems. The basic idea of this new algorithm is to also introduce a competition between the weights in order to allow only one unit to reach the centre of each cluster. The algorithm was tested on multi-band images with different starting weight positions, giving similar results.
Paolo Gamba, Andrea Marazzi, Alessandro Mecocci

A Comparative Study of Topological Feature Maps Versus Conventional Clustering for (Multi-Spectral) Scene Identification in METEOSAT Imagery

A connectionist scheme based on auto-adaptive topological feature maps is compared to conventional cluster analysis for (multi-spectral) scene identification in METEOSAT data. The identification of scenes in (multi-spectral) satellite data is equivalent to the assignment of meaningful labels to spatially coherent image regions. Automated scene detection in METEOSAT data is considered as a data reduction problem with the (optimal) preservation of the spatial coherence of scene region information. A self-organising one-dimensional feature map applied to the so-called segment space of the individual METEOSAT channels is shown to be an appropriate tool for mono-spectral scene identification. It is also shown that the presented connectionist approach for auto-adaptive mono-spectral scene identification has two important advantages compared to conventional cluster analysis, i.e. the number of detected regions is never lower than the number of nodes in the feature map and one obtains a contrast enhancement of the image regions. It is argued however that, despite these properties of feature maps, the use of conventional cluster analysis must be preferred for multi-spectral scene identification because multi-dimensional feature maps become far too slow for practical applications and require a detailed statistical analysis of the multi-spectral segment distribution in order to choose the topological dimension of the feature maps.
P. Boekaerts, E. Nyssen, J. Cornelis

Self Organised Maps: the Combined Utilisation of Feature and Novelty Detectors

With the interest in neural network implementations for real time applications rising continuously, it becomes important to obtain an insight into their performance and to increase the accuracy of the results obtained by them.
During the classification phase of a neural network a new sample is introduced to the network and the maximum response, i.e. the “best match”, is sought. The node with the maximum response then becomes the winner. This method, despite its benefits, provides no indication concerning the degree of match between the new sample and the winning node. Most neural network implementations provide no indication of an input sample that was never “seen” in the past i.e. when nothing similar has ever been used as a sample for training.
To alleviate the above-mentioned problem the Self Organised Feature Detector Map (SOFDM) and the Self Organised Novelty Detector Map (SONDM) were tested. In this paper we present the first results of the combined implementation of the two “conjugate” self organised maps. The SOFDM is a Self Organised Map (SOM) neural network that becomes specifically tuned to the input samples used for training. The SONDM is organised in a way that will result in a small output for “known” patterns, while its output will be very high for “unseen”-novel patterns, thus recognising new patterns.
In an attempt to assess the performance of the networks during their training a measure of disorder is introduced. Based on this measure, different algorithmic approaches are also tested using real data from the National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR).
C. N. Stefanidis, A. P. Cracknell

Generalisation of Neural Network Based Segmentation Results for Classification Purposes

In this paper the automatic post-processing of segmented images is discussed. The segmentation based on local features and neural networks produces often small regions that disturb further analysis. Strategies for the elimination of these small regions are discussed. One approach based on pyramidal hierarchy is implemented. The approach is tested on a land-based cloud classification problem and the results are reported. This simple strategy applied on the cloud classification problem improves the result 20–30 per cent depending on the image. In future continuation of this work it is planned to study how the dynamical expanding context and learning grammars can improve the generalisation of the segmentation and the classification result.
Ari Visa, Markus Peura

Remote Sensing Applications Which may be Addressed by Neural Networks Using Parallel Processing Technology

More instruments per payload, instruments which are increasingly sophisticated, and a fleet of Earth observation satellites will all conspire to flood ground stations throughout the world with an unprecedented quantity of data by the turn of the century.
This deployment of remote sensing hardware is the response of Earth observation agencies world-wide to the challenge of finding out more about the Earth’s natural systems and the stresses placed upon them by human activity.
Acquiring the data however only addresses one part of the challenge. Another equally critical requirement is the need to transform the acquired data into timely, high quality, and cost-effective information that can be used by Government agencies, scientists, and industry to enable better stewardship of the Earth and its resources.
Not only is the quantity and rate of data capture increasing but the trend for performing multi-temporal analyses of, already formidable, datasets is also likely to tax the data-processing facilities of even the best equipped ground stations.
This paper examines how the coupling of High Performance Computer Networks (HPCNs) and the data processing potential of Neural Networks (NNs) may have important consequences for the efficient generation of some Earth observation products and information.
Charles Day

General Discussion

The participants to the COMPARES workshop were asked to consider and reflect upon a set of key questions which relate to the future use of neurocomputing in remote sensing. These questions were as follows:
Why use neural networks in the analysis and interpretation of remote sensing data ? Do they really offer advantages ? Are they really special ?
Classification has been the main application to date but has an impasse been reached ?
what this book tries not to do
Is it possible or necessary to build a “pan-European-classifier” based on a giant modular neural network ?
Is special purpose hardware needed to operationally exploit neural networks in remote sensing ?
Should new/less common network models/architectures be explored ?
Are there any novel applications of neural networks in remote sensing ?
What future level of investment on R&D on neural networks in remote sensing is justified ?
Graeme G. Wilkinson
Weitere Informationen