Tropical Cyclone Center Location in SAR Images Based on Feature Learning and Visual Saliency

Abstract

Synthetic aperture radar (SAR), with its high spatial resolution, large area coverage, day/night imaging capability, and penetrating cloud capability, has been used as an important tool for tropical cyclone monitoring. The accuracy of locating tropical cyclone centers has a large impact on the accuracy of tropical cyclone track prediction. This study focuses on the center location of tropical cyclones in the SAR images. Based on the analysis of the characteristics of the tropical cyclone SAR images, combined with the theory and methods of SAR image segmentation and computer vision, center location methods for both the tropical cyclones with eyes in the SAR images and the tropical cyclones without eyes in the SAR images are presented in this chapter. The main work is as follows: 1, For a tropical cyclone with its eye in the SAR image, The eye area in image appears as black or dark grey area for there being no rain and little wind in the eye area. But the gray level contrast is not always obvious. There may be no complete and clear eye when a tropical cyclone is in the development period or the recession period. The eye area in the tropical cyclone SAR image may appears as light grey area at these periods. So it is necessary to enhance the gray level contrast before image segmentation. Besides, denoising the speckle noise is also necessary for the SAR image processing. A tropical cyclone eye extraction method based on non-local means method and labeled watershed algorithm is given. PPB filter is used to denoise the speckle noise. Then the top-hat transform is used to enhance the contrast. At last the tropical cyclone eye is extracted labeled watershed algorithm. The eye area extracted with this method is computed to compare with the eye area extracted manually. The comparison indicates the accuracy of the extraction accuracy. 2, Generally speaking, the center of the tropical cyclone without its eye is located with template matching method for a single image. The spiral cloud band of the tropical cyclone without its eye is the information can be fully used in the tropical cyclone SAR image. Take the advantage of simple background with little texture information, a center location method of the tropical cyclone without its eye in the SAR image based on feature learning and visual saliency detection is proposed. Spiral cloud bands appear as light and dark spiral structure in the tropical cyclone SAR image, containing rich directional information. Therefore salient region map taking advantage of the gray contrast feature and orientation feature is built. The salient region map makes the spiral cloud bands outstanding and the irrelevant clouds excluded. Then the morphology method is used to extract the spiral bands in the salient region map, the skeleton lines of spiral cloud bands is extracted. At last the tropical cyclone center is estimated with the inflow angle model and the particle swarm optimization algorithm. And the estimation results are compared with the Best Track Data, confirming the validity of the algorithm.