Comparison of local transfer function classifier and radial basis function neural network with and without an exhaustively defined set of classes

Abstract

The local transfer function classifier (LTF‐C) is a new radial basis function (RBF)‐like neural network, but it uses an entirely different learning algorithm, so as to achieve the novel ability of locally partitioning the feature space. This paper investigates LTF‐C and the RBF neural network with reference to land cover classification with and without an exhaustively defined set of classes using Landsat‐5 TM data. Results indicate that LTF‐C achieves higher accuracy, usually with fewer hidden units, than the RBF neural network with an exhaustively defined set of classes. LTF‐C is more stable than the RBF neural network during classifications of the testing set, including the untrained class. Through the setting of post‐classification thresholds on the network's outputs, a well‐trained RBF neural network sometimes gives abnormally high output value for an input pattern which represents the untrained class. Meanwhile, a well‐trained LTF‐C outputs extremely low values all the time under the same circumstances. Therefore, LTF‐C may outperform the RBF neural network in detecting or removing the atypical classes that are excluded from the training set, which maybe useful in situations where only interesting types of land cover are selected in the training set, due to high labour costs or difficulties in defining all classes represented in a study area.

Acknowledgements

The authors are grateful to the anonymous referees for their valuable comments and suggestions. Also, the authors would like to thank Macin Wojnarski for providing us the source code of LTF‐C. This study is financially supported by the 973 projects (2007CB714406) and the 863 projects (2006AA12Z130, 2007AA12Z157).