Research And Application Of Hyperspectral Region Grow Algorithm Combining Unmixing

With the development of offshore oil transportation,oil spill accidents occur more frequently and its pollution to the sea is undoubtedly devastating.After the accidentjhyperspectral remote sensing can give us a strong evidence to estimate the volume of oil spilling by providing the information about the position of oil spill area and the thickness of the oil slick timely.But now,most of the segmentation algorithms is not satisfactory when put in the regional division of oil spilling images.Therefore.it’s necessary to do further research about it.Combined with the hyperspectral technology and oil spilling images characteristics,the contributions based on the existing region grow algorithm are summarized as follows:Firstly,unsupervised endmember extraction method is used to get seeds automatically.And then,we obtain the abundance image which is corresponding to the hyperspectral image by unmixing.So that region grow algorithm is used on the two-dimensional abundance image.Secondly,We proposed three new growth patterns against the noisy influence of the undulating surfaces and the solar flare,namely neighborhood average growth pattern,neighborhood extremes growth pattern and neighborhood average growth pattern without outliers. We compare these three patterns with the orginal growth pattern,then compare the results with threshold segmentation method and cluster segmentation method respectively to verify the validity of the method.Finally,we use the mean variance inside of the boundary curve and the mean gradient along the curve as parameters to build the model which is used to choose the best threshold during region grow process.So we can make sure to get the best segmentation.In this paper,we put Synthetic image,PengLai 19-3C hyperspectral oil image and USA(Indiana) agroforestry Proving Ground AVIRIS hyperspectral image in our experiments for our study.According to the results of the experiment,the proposed method of combine unmixing algorithm and use improved growth patterns is feasible and effective. This method can improve the accuracy of segmentation by reducing the the points misclassified and drained.