Efficient sea–land segmentation using seeds learning and edge directed graph cut

• A supervised method combining seeds learning and graph cut (GC) is proposed to address the problem of sea–land segmentation.
• We propose a multi-feature descriptor to describe the sea and the land, classification results on testing samples demonstrate the effectiveness of the descriptor.
• Superpixel method is used to extract samples and build graph model for GC, which will reduce information redundancy and enhance the local clustering property of neighboring pixels.
• Edge information between neighboring superpixels is incorporated when building the boundary term of GC to reduce the cost of separating superpixels at two sides of the edge into different parts, which can help to avoid under-segmentation for some thin and elongated structures.
• Segmentation results of our method outperform that of state-of-the-art methods in terms of quantitative and visual performance.

Separating sea surface and land areas in an optical remote sensing image is very challenging yet of great importance to the coastline extraction and subsequent inshore and offshore object detection. The state-of-the-art methods often fail when the land and sea areas share complex and similar intensity and texture distributions. In this paper, we propose a graph cut (GC) based supervised method to segment the sea and the land from natural-colored (red–green–blue, RGB) images. Firstly, an image is pre-segmented into superpixels and a graph model with the superpixels as its nodes is constructed. Then each superpixel node is encoded by a multi-feature descriptor, and a probabilistic support vector machine (SVM) is trained for automatic seed selection. These seeds will be used to build the prior model for GC. When modeling boundary term in GC, we incorporate edge information between neighboring superpixels to get finer results for some thin and elongated structures. Experiments on a set of natural-colored images from Google Earth demonstrate that our method outperforms the state-of-the-art methods in terms of quantitative and visual performances.