Hyperspectral Anomaly Detection Based On Nonlinear Kernel Mapping

In the modern development of remote sensing technology,hyperspectral anomaly detection has also been emerged rapidly and has gradually become a research hotspot.Different from the conventional optical imagery,hyperspectral imagery can effectively distinguish subtle spectral differences through its unique feature spectral information.In real applications,it is extremely difficult to obtain prior information.As a result,anomaly detection without any priori information has evident practicability and application value.However,in real anomaly detection applications,along with the growing increase of spectral information,the ever-expanding hyperspectral dataset has also brought tremendous pressure and challenges to data storage,transmission and further processing;on the other hand,there exists large area background interference in small target detection,and it will result in a higher false alarm rate by the direct use of traditional threshold segmentation method.The segmentation thresholds of existing anomaly detection algorithms are normally determined by several experiments,making it difficult to obtain optimal threshold value in real applications.Therefore,how to reduce ground-based data storage and processing workload,improve abnormal target detection accuracy and effectiveness is a pressing issue.Based on the above problems,aiming at the problem of improving detection efficiency,a recursive kernel anomaly detection algorithm based on kernel RX is firstly proposed in this paper;and then an adaptive threshold segmentation for hyperspectral anomaly detection is proposed to solve the problem which the threshold is difficult to choose.The main contents include the following aspects:First of all,in practical applications,fast algorithm for hyperspectral anomaly detection is particularly important,and traditional kernel based RX algorithm has the problem of low efficiency caused by complicated calculation..Concerning this issue,this paper proposes a fast anomaly detection algorithm using recursive kernel method.Different from the traditional kernel method,the data samples used for background estimation should be only those up to the data sample vector currently being processed,and the only storage should include the currently being detected pixel and previous kernel matrix,spectra within the whole detection area are no longer needed for storage,which not only can reduce storage space,but also can realize real-time processing along with data transmission;then the update equation for kernel matrix is built up by deriving recursive causal relationship to avoid reduplicate calculation of high-dimensional space mapping of kernel matrix,in the sense that the processing time is greatly reduced and the algorithm efficiency is largely improved.Simulation results show that the proposed method could results in an accurate detection result with less processing time,so it has a strong practical value.Secondly,there exists large area background interference using the kernel RX anomaly detection algorithm.Concerning this issue,this paper presents a new anomaly detection method based on morphological background suppression,which makes use of morphological filtering to extract large area of the background clutter to remove background interference and improve detection performance.Experimental results show that this method has good ability of large area background interference suppression and abnormal target maintain and it could significantly improve the performance of existing kernel RX algorithm for abnormal small target detection.Finally,it is crucial that sometimes the global threshold segmentation will fail to work for those anomalies existed only in the local area or those anomalies with weak energies.To solve this problem,this paper makes use of an iterative method for adaptive threshold selection,and proposes a local adaptive thresholding method.The grayscale image of detection result is divided into several sub-images,and then the proposed iterative method for adaptive threshold is conducted on each sub-image,respectively.Finally,the decision results for each sub-mage are traversed through the whole image to get the final detection result.Experimental results show that the proposed method could result in a better detection performance as well as an adaptive selection of optimal threshold and is feasible for practical applications.