Multi-feature sea–land segmentation based on pixel-wise learning for optical remote-sensing imagery

Robust sea–land segmentation in optical remote-sensing images is challenging because of the complex sea–land environment and scene diversity. Here, we propose a novel multi-feature sea–land segmentation method via pixel-wise learning for optical remote-sensing images. Multiple features such as greyscale, local statistical information, edge, texture, and structure are first extracted from each pixel in training images and then used to learn a multi-feature sea–land classifier, which transforms the segmentation issue into pixel-wise binary classification problem. In our approach, a new multi-feature sea–land segmentation algorithm is put forward based on the approximation of Newton method. Experiments on Google-Earth, Venezuelan Remote Sensing Satellite-1 (VRSS-1) and Gaofen-1 images demonstrate that the proposed approach yields more robust and accurate sea–land segmentation results.