The Study Of Forensics Identification Method Based On Hyperspectral Image

As one of the important technical means of forensics identification, optical identification has greater advantages than others on its characteristics of nondestructive inspection. But traditional optimal imaging methods have certain requirements on the experiment environment. And data obtained by them is too small to achieve good forensics identification’s result. Different from others, hyperspectral imaging technology may obtain data which has abundant information on both space and spectrum. With the development of science and technology, hyperspectral images’ spatial resolution is getting higher and higher that makes hyperspectral images applied to other areas rather than remote sensing field only.To avoid inaccurate classification problem caused by traditional hyperspectral images’low spatial resolution, we take high spatial resolution hyperspectral images by LCTF. This paper adopted both unsupervised and semi-supervised learning methods to research on forensics identification.An adaptive hyperspectral image clustering method which fuses spatial and spectral information was proposed to get better result of forensics identification. Clustering algorithms, such as K-Means, ISODATA depend too much on parameters to solve forensics identification well. Taking advantage of strong space constraints between pixels of high spatial resolution hyperspectral images, adaptive unsupervised clustering method based on MAP-ML Model and BIC was proposed. The results show that adaptive hyperspectral clustering method based on fusion of space and spectrum not only can select optimal classes’number automatically but also has better clustering accuracy than K-Means and ISODATA algorithms. Furthermore, it has good stability.Besides, labeled sample data we can get is limited in the actual forensics identification applications. In order to maximize the use of them, this paper proposed a semi-supervised learning classification method to improve the accuracy. However, owing to the addition of a priori knowledge semi-supervised learning method proposed in this paper obtains higher classification accuracy and better results than adaptive hyperspectral clustering method.