Demetris Stathakis

How many hidden layers and nodes

The question of how many hidden layers and how many hidden nodes should there be always comes up in any classification task of remotely sensed data using neural networks. Until today there has been no exact solution. A method of shedding some light to this question is presented in this paper. A near‐optimal solution is discovered after searching with a genetic algorithm. A novel fitness function is introduced that concurrently seeks for the most accurate and compact solution. The proposed method is thoroughly compared to many other methods currently in use, including several heuristics and pruning algorithms. The results are encouraging, indicating that it is time to shift our focus from suboptimal practices to efficient search methods, to tune the parameters of neural networks.

Comparison of computational intelligence based classification techniques for remotely sensed optical image classification

Several computational intelligence components, namely neural networks (NNs), fuzzy sets, and genetic algorithms (GAs), have been applied separately or in combination to the process of remotely sensed data classification. By applying computational intelligence, we expect increased accuracy through the use of NNs, optimal NN structure and parameter determination via GAs, and transparency using fuzzy sets is expected. This paper systematically reviews and compares several configurations in the particular context of remote sensing for land cover. In addition, some of the configurations used here, such as NEFCASS and CANFIS, have few previous applications in the field. A comparison of the configurations is achieved by testing the different methods with exactly the same case-study data. A thorough assessment of results is performed by constructing an accuracy matrix for each training and testing data set. The evaluation of different methods is not only based on accuracy but also on compactness, completeness, and consistency. The architecture, produced rule set, and training parameters for the specific classification task are presented. Some comments and directions for future work are given

Investigation of genetic algorithms contribution to feature selection for oil spill detection

Oil spill detection methodologies traditionally use arbitrary selected quantitative and qualitative statistical features (e.g. area, perimeter, complexity) for classifying dark objects on SAR images to oil spills or look‐alike phenomena. In our previous work genetic algorithms in synergy with neural networks were used to suggest the best feature combination maximizing the discrimination of oil spills and look‐alike phenomena. In the present work, a detailed examination of robustness of the proposed combination of features is given. The method is unique, as it searches though a large number of combinations derived from the initial 25 features. The results show that a combination of 10 features yields the most accurate results. Based on a dataset consisting of 69 oil spills and 90 look‐alikes, classification accuracies of 85.3% for oil spills and in 84.4% for look‐alikes are achieved.

Neural networks as a tool for constructing continuous NDVI time series from AVHRR and MODIS

The long term Advanced Very High Resolution Radiometer (AVHRR)‐Normalized Difference Vegetation Index (NDVI) record provides a critical historical perspective on vegetation dynamics necessary for global change research. Despite the proliferation of new sources of global, moderate resolution vegetation datasets, the remote sensing community is still struggling to create datasets derived from multiple sensors that allow the simultaneous use of spectral vegetation for time series analysis. To overcome the non‐stationary aspect of NDVI, we use an artificial neural network (ANN) to map the NDVI indices from AVHRR to those from MODIS using atmospheric, surface type and sensor‐specific inputs to account for the differences between the sensors. The NDVI dynamics and range of MODIS NDVI data at 1° is matched and extended through the AVHRR record. Four years of overlap between the two sensors is used to train a neural network to remove atmospheric and sensor specific effects on the AVHRR NDVI. In this paper, we present the resulting continuous dataset, its relationship to MODIS data, and a validation of the product.

Satellite image classification using granular neural networks

The increased synergy between neural networks (NN) and fuzzy sets has led to the introduction of granular neural networks (GNNs) that operate on granules of information, rather than information itself. The fact that processing is done on a conceptual rather than on a numerical level, combined with the representation of granules using linguistic terms, results in increased interpretability. This is the actual benefit, and not increased accuracy, gained by GNNs. The constraints used to implement the GNN are such that accuracy degradation should not be surprising. Having said that, it is well known that simple structured NNs tend to be less prone to over‐fitting the training data set, maintaining the ability to generalize and more accurately classify previously unseen data. Standard NNs are frequently found to be accurate but difficult to explain, hence they are often associated with the black box syndrome. Because in GNNs the operation is carried out at a conceptual level, the components have unambiguous meaning, revealing how classification decisions are formed. In this paper, the interpretability of GNNs is exploited using a satellite image classification problem. We examine how land use classification using both spectral and non‐spectral information is expressed in GNN terms. One further contribution of this paper is the use of specific symbolization of the network components to easily establish causality relationships.

Large-scale feature selection using evolved neural networks

In this paper computational intelligence, referring here to the synergy of neural networks and genetic algorithms, is deployed in order to determine a near-optimal neural network for the classification of dark formations in oil spills and look-alikes. Optimality is sought in the framework of a multi-objective problem, i.e. the minimization of input features used and, at the same time, the maximization of overall testing classification accuracy. The proposed method consists of two concurrent actions. The first is the identification of the subset of features that results in the highest classification accuracy on the testing data set i.e. feature selection. The second parallel process is the search for the neural network topology, in terms of number of nodes in the hidden layer, which is able to yield optimal results with respect to the selected subset of features. The results show that the proposed method, i.e. concurrently evolving features and neural network topology, yields superior classification accuracy compared to sequential floating forward selection as well as to using all features together. The accuracy matrix is deployed to show the generalization capacity of the discovered neural network topology on the evolved sub-set of features.

Granular neural networks for land use classification

Granulation of information is a new way to describe the increased complexity of natural phenomena. The lack of clear borders in nature calls for a more efficient way to process such data. Land use both in general but also as perceived in satellite images is a typical example of data that are inherently not clearly delimited. A granular neural network (GNN) approach is used here to facilitate land use classification. The GNN model used combines membership functions of spectral as well as non-spectral spatial information to produce land use categories. Spectral information refers to IRS satellite image bands and non-spectral data are here of topographic nature, namely slope, aspect and elevation. The processing is done through a standard neural network trained by back-propagation learning algorithm. A thorough presentation of the results is given in order to evaluate the merits of this method.

Prediction of crop yields with the use of neural networks

The possibilities of the combined use of neural networks and fuzzy set theory in the form of constructing a so-called fuzzy neural network (FNN) or granular neural network (GNN) [1] for predicting crop yields in the Rostov oblast and Krasnodar and Stavropol krais are examined. The results of modeling plant growth on the basis of the CGMS simulation model as well as the values of the vegetation index NDVI, calculated from the SPOT VEGETATION satellite data, are the input parameters. As a result of training the neural network, the accuracy of predicting yields is on average about 75%.

Efficient segmentation of urban areas by the VIBI

Urban populations are expanding rapidly and so are cities. Remote sensing offers a convenient means of monitoring this expansion as it covers a period of 40 years in the case of the LANDSAT satellite. In some parts of the globe, this is probably the only viable means of monitoring due to the lack of other types of data. In order to monitor expansion, first, urban land has to be separated from other land-cover types. Although this can be done by standard classification processes, it is much more efficient to establish an urban index (UI) analogous to the widely used normalized difference vegetation index (NDVI) for vegetation. Existing efforts to establish such a UI are reviewed and compared in a common context. Following this, a novel, more efficient UI is introduced. The calculation of the new index is straightforward, based on combining the NDVI with the normalized difference built-up index. The results are promising as the index can efficiently segment urban areas, even in the presence of excessive bare land. The proposed method is evaluated on two test sites selected in different LANDSAT scenes. The new index is valid only for sensors with the same bands as those of LANDSAT.

Examining urban sprawl in Europe using spatial metrics

Urbanisation is a global phenomenon with an important impact on the quality of human life. Europe has been widely affected by urbanisation. One of the main characteristics of urban growth is sprawl, a negative form of urban expansion, which affects large cities and most types of urban landscapes. Spatial indicators are applied to CORINE Urban Morphological Zones (UMZ) changes in order to measure urban sprawl between 1990–2000 and 2000–2006 in 24 European countries. The indicators calculate urban morphological properties such as shape, aggregation, compactness and dispersion. The results revealed that the urban areas (UMZ) increased by 146% during 1990–2006 and the urbanisation becomes more circle-shaped and less complex where mostly sprawl occurs. Moreover, urban form becomes less clumped or aggregated. Therefore, due to accelerating rates of urban sprawl, European urban planning should intensify appropriate initiatives to avoid negative impacts on human life.