Small anomaly detection in ocean evironment is an important problem in airborne remote sensing image processing, especially in hyperspectral data. Traditional algorithms solve this problem by finding the pixels have different appearance pattern with the background. However, these algorithm are not suitable for real-time applications. In this paper, we propose to learn the hyperspectral model of the seawater and localize the targets whose spectral feature do not well fit the trained model. This algorithm only uses historical information and is suitable to be used on airborne line-scanning data. Since hyperspectral property of ocean water is relatively stable, we use Gaussian mixture model to encode the statistical features of the background. Experimental results demonstrated that the proposed algorithm significantly improves processing efficiency in comparison with conventional methods, and maintains high accuracy with regard to other methods.