A Novel and High Performance Change Detection Algorithm for Hyperspectral Images

Large format data such as hyperspectral images contain a lot of information about various targets. Subpixel detection is now achievable. As a result, these images are becoming popular in military surveillance reconnaissance operations. However, one serious limitation of hyperspectral images is that many constituent members (materials) may be present in a single pixel, making accurate target detection an extremely challenging task. We propose a novel and high performance approach to change detection using hyperspectral images. The novel approach consists of several key components. First, we propose to apply an unsupervised algorithm to extract signatures (endmembers) from hyperspectral images. This approach achieves accurate signature extraction performance from highly mixed pixels as compared to conventional methods. It does not require the presence of pure pixels in a given hyperspectral image. Second, we propose a nonlinear gradient descent based method to perform abundance estimation, which is related to change/target detection. Our approach is based on maximum entropy and achieves robustness to noise. It is also applicable to nonlinear mixtures. BENEFIT: The proposed technology will be very useful for both military and commercial applications. Here we briefly highlight some potential markets where the proposed algorithms will be applicable. Many military (DoD) applications including reconnaissance and surveillance, homeland security, perimeter defense, etc. will benefit from this technology. In addition, Lockheed Martin, Raytheon, GE, MITRE, are also potential customers for this technology.  The market for military applications is quite large. We expect the market size will be at 20 million dollars over the next decade.)