Feature Level Information Fusion Based on Neural Network and Its Application to Airport Recognition

This paper deals with the airport recognition from optical remote sensing images under the framework of feature level information fusion by using BP neural network. Automatically extracting airport from remote sensing imagery is one of research hots which has being attracted much research interests both from the computer vision community and remote sensing application field. Due to its uniqueness in geometric shape, shape property analysis on airport and recognition based on shape parameters or directly on shape recognition is the intutitive approch in order to recognize airport. However, though straitfordward in method, recognition based on shape analysis only can cause unavoidalbe mistakes when there are objects similar in shape around the airport such as express ways.Thanks to the multispectral property of most commecial remote sensing images, this paper expound an approach for airport recognition based on multi feature fusion, namely, shape parameter, spectral signature and texture feature of an image. Since all these features are different in unit and order of magnitude, it is difficult to combine these features in commonly used algorithms, such as Bayesian inference. Error Back Propagation Neural Network is used to fuse all these features. Two revised version of BP network is used in order of improving the convergence speed during training and recognizing. Real TM images are used for case study, showing the effectiveness of our approach.