A Network Technique Based Feature Extraction Method For Remote Sensing Images

摘要

For hyperspectral remote sensing, the dataset usually contains hundreds of spectral images, which generates a rather large amount of data. To reduce the computational complexity of image analysis, feature extraction is often adapted. This paper presents a new approach for unsupervised feature extraction by transforming the hyperspectral dataset into complex networks. The networks' statistical topological properties are investigated to evaluate remote sensing image features. The objective of the method is to find the images which can form the most representative network formation. This is a completely new criterion for feature selection. Meanwhile, the proposed technique has both an explicit physical meaning. Experimental results demonstrate that the method achieves better results with respect to traditional methods.