UNDERWATER LINEAR FEATURE EXTRACTION WITH MULTISPECTRAL BAND IMAGES: AN EVALUATION WITH LEVEL-SET METHOD IN DONGSHAATOLL AND ZENGMU SHOAL

Abstract

Optical remote sensing satellite images are a useful and convenient source to provide underwater features, particularly for shallow water areas because light, dependent on wavelength, has the capability to penetrate water. In this study, the information richness of underwater features is investigated for each spectral band of the optical images, and also several derived bands. This assessment is performed with the level-set method for segmentation. Two cases are analyzed in this study. The first study site is the Dongsha atoll, which is composed of Dongsha island, lagoon, and surrounding reefs. The water depth ranges from zero to less than 3 m at the outer ring and down to a depth of 20 m in the lagoon. The images were acquired with WorldView-2 in October 2013 and covered the entire atoll. The second study site is Zengmu shoal, an underwater feature. The image used is a scene acquired with Landsat 8. These images demonstrate high water clarity in both sites. For the Dongsha atoll, both the reflectance of each spectral band, the NDWI, and bands processed with Principle Component Transformation (PCT) are analyzed. The assessment is made based on the number of segments identified. The more segments identified, subsequently the more information, we assume, is provided. In order to remove those caused by noise, only the segments larger than 100 m(2) were counted. Based on this, PCT band 1 performs the best, and followed by green, yellow, coastal, blue, red, and fewer features from red-edge NIR and NIR2 bands when the objects in the scene are completely submerged underwater. For the Zengmu shoal, the boundary of the object identified is used for the assessment. The one closest to the manually. digitized imaged boundary would be recognized as having the best performance. Among the spectral bands, coastal/aerosol (CA) and blue perform the best. The four bands, coastal, blue, green, and red, are projected with PCT. The boundary resulting from the first principle component resembles most the one identified by a human operator on a QuickBird image.