The Study Of Hyperspectral Oil Spill Image Segmentation Based On Active Contour

In recent years, marine oil spill accidents occur frequently, resulting in a large number of oil spills, so that it brings serious harm to the marine environment and organisms. After the oil spill accident, to extract the oil spill boundary from the oil spill remote sensing image quickly and accurately, and then to obtain the oil spill area and the quantity of oil spill is the key link of the oil spill monitoring, and it is also an important research topic. Remote sensing technology is one of the important means of Marine oil spill surveillance, and as a frontier technology in the field of remote sensing hyper-spectral remote sensing, with its unique advantages, becomes a new bright spot for the monitoring of oil spill.Active contour model, compared with other segmentation algorithms, has a main advantage that no matter how the image quality is, it always can obtain smooth and closed boundary of target area. Chan and Vese proposed active contour without edges, namely CV model. The model has a series of advantages such as it does not depend on gradient information, can segment the image which has vague border or contains noise, is not sensitive to the initial curve and easy to implement and so on, therefore it has been used widely. While as a result of hyperspectral oil spill image with fuzzy boundary, little spectrum characteristic difference between oil and water, intensity inhomogeneities and a lot of bright spots and shadows noise, CV model can’t achieve good segmentation effect. In this paper, the CV model is improvement in the following three aspects:(1) By introducing the basic thought of Fisher criterion to improvement of the fitting term of CV model, the improved model can obtain better segmentation effect, at the same time, the model can automatically adjust the ratio between the coefficient of the length term and the coefficient of the fitting term, so that the model is more stable to the changes of parameters.(2) Using spectral angle to construct an edge detection function and building a new length term by introducing the edge detection function, the modified model can make the contour curve to stop at the boundaries of the target area more accurately and stably.(3) Combining endmember extraction algorithm to further improve the CV model, the improved model can segment the particular object of interest from the image which has many kinds of objects.At last, the experimental results on synthetic and real hyper-spectral images have shown that the proposed model obtains more satisfactory segmentation compared to the CV model and other methods, which have proved the feasibility, validity and accuracy of the proposed model.