Hyperspectral Anomaly Detection Based On Isolation Forest Algorithm

Hyperspectral images have hundreds or even thousands of spectral bands,which provide abundant spectral information for each pixel.This advantage of hyperspectral image can distinguish the target from the background more effectively,however,anomaly detection without prior knowledge is still a big challenge for hyperspectral application.Different from the existing anomaly detection methods,iForest(isolation forest)assumes that the abnormal target is more likely to be isolated than the background,and uses the idea of feature integration to realize anomnaly detection.Therefore,on the basis of in-depth study of iForest anomaly detection algorithm,this paper fully considers the characteristics of hyperspectral image data,and adopts iForest for hyperspectral anomaly detection.The main research contents are as follows:(1)iForest randomly selects features from high-dimensional data.However,due to the high correlation among hyperspectral bands,it is very easy to miss the bands with high information content.Therefore,this paper proposes a method of hyperspectral anomaly detection based on isolation forest with band clustering.This method firstly uses the clustering method to obtain several clusters,and then randomly selects one cluster as the candidate set of features,so as to optimize the construction of isolated tree in iForest method.The experimental results show that this method can make full use of the spectral information in the image and improve the detection ability of abnormal targets.(2)The complexity of background information in hyperspectral remote sensing images is generally high,which is not conducive to iForest algorithm’s segmentation of input features.Therefore,this paper proposes an iForest hyperspectral anomaly detection method based on background orthogonal-subspace enhancement.Firstly,the original data is projected into the orthogonal subspace of the background to increase the contrast between the potential abnormal target and the background.iForest is then applied to the projected data.Among the four groups of hyperspectral data experiments,the proposed algorithm has the best detection results,in which the AUC value of airportl data is increased by 9.4%compared with that of iForest algorithm.(3)Classical iForest method uses pixels that are selected in the whole image to construct isolated trees,but this global anomaly detection method would generate a large number of false alarms in the results.Both of the above two methods adopt the idea of classical iForest algorithm,and there would also be some false alarms in the detection results.Therefore,this paper proposes two local methods that combine with the difference in the spectrum of abnormal targets and the scarcity of spatial distribution,named localized iForest refinement processing and spatially weighted anomaly expression,to adapt to different circumstances.The experimental results show that the two local iForest anomaly detection algorithms can highlight the abnormal target and better suppress the background information in the image.