A hyperspectral small target detection method based on outlier detection

In the present paper, a new method was discussed, which used outlier detection algorithms of spectral analytical technology to tell the small target from background. First, continuum removal and standard normal variate were employed to pretreat the AVIRIS remotely-sensed data. It could be regarded that continuum is the absorption of background, and the characteristic absorptions are superposed on the continuum. So the continuum removal is frequently applied to remotely-sensed hyperspectral data to eliminate the contribution of background absorption and separate the characteristic absorption of concerned objects from background. SNV corrects each spectrum by subtracting the mean and dividing by the standard deviation for that spectrum. After the pretreatment, spectral angle mapping was used to reduce the dimension of hyperspectral data. Since this mapping process calculated the similarity between the spectra of each pixel and the average spectrum, no prior information such as standard spectrum library is required. And then, Mahalanobis distances were calculated. As Mahalanobis distance shows the extent to which samples deviate from the total population, the points whose Mahalanobis distance are larger than adaptive threshold were regarded as small target. The adaptive threshold was determined by data mean value and maximum value. Applying the algorithm above to a set of AVIRIS remotely-sensed hyperspectral data which was free downloaded from the NASA official website, small target on the concerned area was picked out correctly. In addition, the above mentioned took about 1/8 time as much as the traditional Mahalanobis distance method without prior reducing dimension of hyperspectral data. Compared with traditional algorithm, no prior information is needed, and less calculating work and time is required. Still, it has got a satisfying accuracy.