Research On Anomaly Detection Algorithm With Window-Based Method

Target detection is one of the most important applications of hyperspectral remote sensing.But in practice,anomaly detection is often used to face the problem of lacking priori knowledge of targets instead of target detection.Which algorithms based on probability and statistical model,this kind of detectors have two defects in the application,these detectors use the probability density functions of the multivariate normal distribution to measure the probability of the test pixel to be part of the background.Due to the existence of noisy and other anomalous pixels,the estimated covariance matrix and the mean vector globally as a form of background representation may not be accurate.The higher spectral resolution made the hyperspectral images with a lower spatial resolution,which made the targets as small targets with low outcrop probability in hyperspectral images only with a few pixels or less than one pixel.Classical anomaly detectors with low precision could not deal the problem of high precision requirement applications as rescue after the earthquake,flight monitoring,vessels searching,air crash rescuing,military scouting,etc.In this paper,to deal with the problem that detector using global method can mixed targets' information with the background characteristics,we analyzing the principal of the classical algorithms,combined classical algorithms with dual-window method,and propose the algorithms based on dual-window and similarity measurement model,provided more precise algorithms for high precision anomaly detection.Through this research,we came up following conclusions:(1)The key problems of target detection are systematically investigated in this paper,aiming at the requirement using anomaly detection about small targets with low outcrop probability.Based on probability and statistical model,compared with the detection precision and analyzed the limitations of classical algorithms,we separated the under test pixels from the background by using dual-window method,transformed the global detection into localization detection,reduced the noise from background characteristics by using the principal component analysis method,improved the classic algorithms as DW-LRX,DW-UTD,DW-LPTD;Combined the dual-window method with similarity measurement model,we evaluate the similarity between the pixel under test and background by using the index like ED,SAM,SID,SCM,and proposed algorithms named DW-ED,DW-SAM,DW-SID,DW-SCM;AVIRIS image standard data from aircraft is used to verify the algorithms accuracy,and results shows modified algorithms can reduce the false alarm more efficient,and higher precision than classic algorithms.(2)In this paper,16 different sizes dual-window are designed to verify the algorithm sensibility of the window size,the result shows the modified algorithms which based on similarity measurement model is more stable than the algorithms based on probability and statistical model.The algorithm which dual-window has a bigger IWR and a smaller OWR can have a higher detection accuracy.5 gradients noise was used to degrade the SNR of AVIRIS image,in order to verify the algorithms sensibility of the noise,result shows that the modified algorithms which based on similarity measurement model has a greater noise resistance;Finally,we testified the algorithm accuracy in application that small targets with low outcrop probability by using hyperspectral data from GF-5 satellite,results shows that modified algorithms greater than classic algorithms,DW-LRX detector has the highest accuracy,DW-SAM and DW-SCM also perform well,reducing the false alarm rate effectively.This paper provide specific algorithms for anomaly detection of small targets with low outcrop probability.