Based On Active Learning Anomaly Detection Hyperspectral Remote Sensing Image

Hyperspectral imagery is a new remote sensing technology which rose in80s of last century, it can provide more detailed information on surface features, it makes the information which cannot be detected by using traditional multi-spectral remote sensing detected. Hyperspectral image target recognition is divided into target detection and anomaly detection, compared to target detection, anomaly detection algorithms can detect objects which have significant differences in spectra from their surrounding background without knowing any prior information, they have good practicability, therefore hyperspectral anomaly detection has been a hot topic in these years.Active learning based Support Vector Data Description(SVDD) anomaly detection algorithm is proposed, aiming at the large amount of calculation problem of traditional SVDD algorithm which is caused by randomly select the training samples in background modeling, this paper introduces active learning ideas of machine learning to optimize training sample selection process, actively choose the most useful samples for building anomaly detection algorithm. Respectively use the algorithm of this paper and the traditional SVDD algorithm to conduct simulation experiments, the result shows that the active learning based SVDD algorithm can make the running time greatly reduced.This paper proposed active learning SVDD algorithm that combines with neighborhood clustering segmentation, which is to take a further research of the function of active learning in reducing the anomaly detection algorithm’s computation complexity. Use neighborhood clustering segmentation to process the remote sensing image and obtain the potential anomaly pixels before proceeding anomaly detection, and then detect the potential anomaly pixels with SVDD algorithm, thus can decrease the calculation complexity. In addition, introduce active learning into SVDD algorithm that combines with neighborhood clustering segmentation, simulation experiment and the experiment which conducted using real Airborne Visible Infrared Imaging Spectrometer(AVIRIS) data shows, when comparing with the SVDD algorithm that combines with neighborhood clustering segmentation, the active learning based SVDD algorithm that combines with neighborhood clustering segmentation can greatly reduce the calculation amount.