An investigation of the design and use of feed-forward artificial neural networks in the classification of remotely sensed images
Artificial neural networks (ANNs) have attracted the attention of researchers in many fields, and have been used to solve a wide range of problems. In the field of remote sensing they have been used in a variety of applications, including land cover mapping, image compression, geological mapping and meteorological image classification, and have generally proved to be more powerful than conventional statistical classifiers, especially when training data are limited and the data in each class are not normally distributed. The use of ANNs requires some critical decisions on the part of the user. These decisions, which are mainly concerned with the determinations of the components of the network structure and the parameters defined for the learning algorithm, can significantly affect the accuracy of the resulting classification. Although there are some discussions in the literature regarding the issues that affect network performance, there is no standard method or approach that is universally accepted to determine the optimum values of these parameters for a particular problem. In this thesis, a feed-forward network structure that learns the characteristics of the training data through the backpropagation learning algorithm is employed to classify land cover features using multispectral, multitemporal, and multisensory image data. The thesis starts with a review and discussion of general principles of classification and the use of artificial neural networks. Special emphasis is put on the issue of feature selection, due to the availability of hyperspectral image data from recent sensors. The primary aims of this research are to comprehensively investigate the impact of the choice of network architecture and initial parameter estimates, and to compare a number of heuristics developed by researchers. The most effective heuristics are identified on the basis of a large number of experiments employing two real-world datasets, and the superiority of the optimum settings using the 'best' heuristics is then validated using an independent dataset. The results are found to be promising in terms of ease of design and use of ANNs, and in producing considerably higher classification accuracies than either the maximum likelihood or neural network classifiers constructed using ad hoc design and implementation strategies. A number of conclusions are drawn and later used to generate a comprehensive set of guidelines that will facilitate the process of design and use of artificial neural networks in remote sensing image classification. This study also explores the use of visualisation techniques in understanding the behaviour of artificial neural networks and the results produced by them. A number of visual analysis techniques are employed to examine the internal characteristics of the training data. For this purpose, a toolkit allowing the analyst to perform a variety of visualisation and analysis procedures was created using the MATLAB software package, and is available in the accompanying CD-ROM. This package was developed during the course of this research, and contains the tools used during the investigations reported in this thesis. The contribution to knowledge of the research work reported in this thesis lies in the identification of optimal strategies for the use of ANNs in land cover classifications based on remotely sensed data. Further contributions include an indepth analysis of feature selection methods for use with high-dimensional datasets, and the production of a MATLAB toolkit that implements the methods used in this study.