基于改进RX增量学习的高光谱图像异常检测

Hyperspectral Image Anomaly Detection Based on Improved RX Incremental Learning

高光谱图像的异常检测是星上处理中重要的研究内容之一,提出了一种在传统RX算法基础上结合增量学习和层级化的高光谱图像异常检测方法。采用增量学习,当生成新的协方差矩阵时不需要计算所有样本的协方差矩阵即可对检测器模型进行更新,避免了数据的重复计算和逆矩阵的求解。利用层级化方法有效地抑制背景,提取目标光谱,增强了检测器的性能。实验结果表明:相较于SAM算法和传统RX算法,所提算法检测概率最高,其检测结果与地面目标最为接近;所提算法的计算复杂度得到了数量级的减弱,与SAM算法相比,运行时间缩短了0.215 s,因此具有更高的检测速度,占用更少的星上资源,优于传统的RX算法。

Anomaly detection of hyperspectral image is one of the important research contents in processing onboard satellite.Based on the traditional RX algorithma hyperspectral image anomaly detection method is proposed by use of incremental learning and the hierarchical method.Incremental learning is used to update the detector model.When generating a new covariance matrixthere is no need to calculate the covariance matrix of all sampleswhich avoids repeated data calculating and inverse matrix solving.The hierarchical method is used to suppress the background and preserve target spectrumwhich effectively improves the performance of hyperspectral image target detector.The experimental results show that:1)Compared with SAM algorithm and the traditional RX algorithmthis algorithm has the highest detection probabilityand its detection result is the closest to the ground target;and 2)The computation complexity of this algorithm is reduced by an order of magnitudethe running time is reduced by 0.215 s compared with SAM algorithm.Thereforethe anomaly detection algorithm proposed here has higher detection speed and occupies less onboard resourceswhich is superior to the traditional RX algorithm.