Research On Anomaly Detection Methods And Realization In Hyperspectral Image

With the development of astrogation and electronics, modern remote sensing technology plays an important role in information gathering. As one of the most significant techniques in observation of the earth in the latest ten years of last century, hyperspectral imaging is becoming a focus in information processing fields at home and abroad. The performance of high spectral resolution makes it suitable for the detection of human-made targets surrounded by natural environment background. Therefore, it becomes more and more popular. However, high dimension and large data amount of hyperspectral images bring trouble to target detection. What's more, traditional segmentation and detection algorithms for single images are inefficient for hyperspectral images. In this application background, the dissertation researches in several aspects as following.Firstly, dimensionality reduction methods are researched. There are two kinds of dimensionality reduction methods, Band Selection and Feature Extraction. Principal Component Analysis (PCA), Segmented Pincipal Components Transform (SPCT) ,band selection based on Genetic Algorithm (GA) and high order component are introduced here, which are all effective in for dimensionality reduction. Band selection methods are more proper for hardware realization and the last method is chosen for dimensionality reduction of hyperspectral image processed by the hardware platform.Secondly, RX, GMRF and SEM, three representational algorithms of anomaly detection are studied. The first two are local anomaly detection algorithms and the last one belongs to global anomaly detection algorithms. They are all be chosen as one of the processing algorithm candidates for the U.S. Defense Advanced Research Projects Agency's (DARPA) Adaptive Spectral Reconnaissance Program (ASRP). The computer simulations are conducted on real hyperspectral data (AVIRIS). Trough comparing the performance and computation complexity of these detectors, RX algorithm is chosen for hardware design.Finally, according to parameters of hyperspectral sensor existing in our...