Land cover mapping of one specific protected habitat under the requirements of the European Union Habitats Directive using remote sensing
Habitat loss is considered to be one of the greatest challenges currently facing society. An important part of the European Union's (EU) response to this problem is the Habitats Directive, the main aim of which is to protect biodiversity through the conservation and protection of natural habitats. Consequently, accurate mapping of specific habitats is of high importance in order to monitor ecosystem changes. Within this context, remote sensing has enormous potential as a source of land cover information which has been used widely for land cover mapping. In many instances this land cover mapping is only concerned with one particular habitat. However, in standard classification analysis, training data for all of the land cover classes contained in the area is typically required which means a wasteful use of resources. This thesis aims to address these issues by investigating advanced classification methods that focus on the accurate mapping of one specific protected habitat. The data used for this purpose is a Landsat ETM+ image of East Anglia, UK, acquired in June 2000 and ground truth data in the form of aerial photography. The habitat of interest chosen for this investigation is fens, a habitat protected by the EU Habitats Directive, whose diverse and dynamic nature is a particular challenge for its mapping and monitoring. A second protected habitat, saltmarshes, will be used for comparison purposes in order to determine any bias within the results. The methods considered to map the selected habitat consist of binary classifiers and one class classifiers. The binary classifiers chosen were Support Vector Machines (SVMs) and Decision Trees (DTs) which are two new methods very recently applied to land cover classification and remote sensing research. They are still very much the focus of current research for multiclass classification. In this thesis they are used in its binary form to classify the class of interest against. all the other classes. Both classifiers perform very well when compared with a classic parametric Maximum Likelihood Classification (MLC). Narrowing down the idea of classifying just one habitat of interest, one-class classifiers are put to the test. They have been explored in pattern recognition research but not yet within remote sensing image classification and land cover mapping. Specifically, the Support Vector Data Description (SYDD) classifier is considered particularly suitable for land cover classification as it is based upon the basis of SVMs which have already been applied in this area with success. When the results of the SYDD classification are compared against those obtained by the other classifiers these show an improvement in overall classification and a reduction in the errors of commission. Furthermore, another method is also put to the test to improve the accuracy of the classification of the class of interest. This method is the ensemble of classifiers, which in many research studies within pattern recognition has proven to improve accuracy of single classifiers. The results in this thesis also show an improvement in accuracy, although further investigation is needed. In conclusion, DT, SVM and SYDD classification methods offer clear advantages over standard classification analysis when concentrating on the classification and mapping of a particular habitat. All three classifiers obtained higher accuracies than the ML classifier with the use of significantly less training data. Furthermore, in the case of one-class classification only data from the class of interest was needed. Also in both binary and one-class classification approaches the attention was focused on separating the class of interest from all the other classes and therefore training efficiency was bigger than in a standard multiclass classification where efforts are directly to achieve a high overall accuracy. All three methods were found to be highly suitable for classifying and mapping a specific habitat and its application should be considered in future work involving the accurate mapping of protected habitats.