Remote Sensing Image Classification Based on a Modified Self-organizing Neural Network with a Priori Knowledge

Abstract

To improve the accuracy of remote sensing image classification based on a self-organizing competitive neural network, this paper firstly uses principal component analysis to reduce redundancy of the multi-spectral remote sensing image data, and then takes the earth surface structure information in horizontal and vertical directions of the target area as a prior knowledge. The self-organizing competitive neural network is modified to contain both structured and unstructured methods. A classifier based on this network, which has been trained by sample data, classifies the remote sensing data from the Landsat TM satellite. The classification results are compared with that from the maximum likelihood estimation classification. The experiment shows that the self-organizing competitive neural network method can improve the accuracy of classification in complex earth surface regions. The overall accuracy and Kappa coefficient are 89.1 % and 0.873, respectively, which outperform the maximum likelihood method by 18.5 % and 0.227. This result illustrates that the proposed method is much better than the maximum likelihood method

Keywords

• Earth surface structure information
• Self-organizing competitive neural network
• Remote sensing image classification
• Maximum likelihood method