A Novel Pixel/Subpixel Target Detection Method for Hyperspectral Image

Target detection is always one of the most important issues in hyperspectral image (HSI) processing. Traditional methods for hyperspectral image target detection always face the problem of lack of information. The training set might be limited in number and incomplete. Sometimes, we even do not have the training set, and the only known prior information is a reference target spectral library. Thus, traditional machine learning approaches cannot perform well in these situations due to the lack of information. This paper puts forward an unsupervised method for detecting pixel/subpixel targets in HSI. In this paper, data field theory is first introduced into the data modeling. Combining with the endmember extract technique, the samples in the high-dimension spectral space are mapped into a low-dimension feature space. Then, an unsupervised learning method is applied to detect the pixel/subpixel targets based on the target properties. According to the experimental results, our method can effectively detect pixel/subpixel targets in HIS with low false alarm rate