Search

International audience; Land-sea segmentation is one of the most challenging tasks in remote sensing. It is of great importance for many applications like coastline extraction and near shore ship detection. To facilitate future land-sea segmentation related research, we present, in this paper, an experimental study conducted on five color spaces (RGB, XYZ, La∗b∗, HSV and YCbCr) in order to show the best-performing color space for land-sea segmentation. The test is carried out on 50 images. The average segmentation results are given. The results show that the luminance-chrominance space (YCbCr, La∗b∗, HSV) are more appropriated than primary space (RGB, XYZ), in general. It is particularly noted that the perceptual color space HSV allows better results in land-sea segmentation application. © 2018 IEEE.

Coastline detection is a very challenging task in optical remote sensing. However the majority of commonly used methods have been developed for low to medium resolution without specification of the key indicator that is used. In this paper, we propose a new approach for very high resolution images using a specific indicator. First, a pre-processing step is carried out to convert images into the optimal colour space (HSV). Then, wavelet decomposition is used to extract different colour and texture features. These colour and texture features are then used for Fusion of Over Segmentation (FOOS) based clustering to have the distinctive natural classes of the littoral. Among these classes are waves, dry sand, wet sand, sea and land. We choose the mean level of high tide water, the interface between dry sand and wet sand, as a coastline indicator. To find this limit, we use a Distance Regularization Level Set Evolution (DRLSE), which automatically evolves towards the desired sea-land border. The result obtained is then compared with a ground truth. Experimental results prove that the proposed method is an efficient coastline detection process in terms of quantitative and visual performances.