Land Cover Classification of Landsat Images Using Problem–Adapted Artificial Neural Networks

Artificial Neural Networks (ANN) have gained increasing popularity as an alternative to statistical methods for classification of Remote Sensing Data. Their superiority to some of the classical statistical methods has been shown in the literature. Therefore, ANNs are commonly used for segmentation and classification purposes. In this paper we address a land cover classification problem using multi–spectral Landsat Thematic Mapper (TM) data employing ANNs. We concentrate on the search for the problem–adapted network topology and the appropriate number of training epochs for Multi–Layer Feed–Forward ANNs. To prevent the ANN from overfitting the number of training epochs is essential. For the automatic generation of problem–adapted topologies a method based on Genetic Algorithms (GA) is employed. With this approach populations of ANNs are generated, trained, and evaluated for the land–cover classification task. Individuals which solve the given task well receive a high fitness value and are selected to form the next generation. A system performing this artificial evolution of ANNs the netGEN system has been implemented by the authors. Due to its computational complexity, netGEN has been designed as a parallel system, i.e. the learning phase of the ANNs is distributed among an arbitrary amount of workstations within a heterogenous cluster. Results for the chosen test site “Nationalpark Hohe Tauern”, where Landsat TM data, a Digital Elevation Model (DEM), and ground–truth data taken from ground measurements are available, are given.