Separating an optical remote sensing image into sea and landareas is very challenging yet of great importance to the coast-line extraction and subsequent object detection. In this paper,we propose a hierarchical region merging approach to auto-matically extract the sea area and employ edge directed graphcut (GC) to accomplish the final segmentation. Firstly, an im-age is segmented into superpixels and a graph-based mergingmethod is employed to extract the maximum area of sea re-gion (MASR). Then the non-connected sea regions are identi-fied by measuring the distance between their superpixels andthe MASR. When modelling the pairwise term in GC, we in-corporate edge information between neighboring superpixelsto reduce under–segmentation. Experimental results on a setof challenging images demonstrate the effectiveness of ourmethod by comparing it with the state-of-the-art approaches. Statistical Region Merging Segmentation
For optical remote sensing imagery, sea-land segmentation isaimed to separate the sea from the land exactly. The seg-mentation result is of great importance to the coastline ex-traction and subsequent ship detection tasks. Exist-ing methods for differentiating between sea and land aremainly based on threshold selection and edge detection, suchas Otsu method and gradient-based method .In multispectral images, normalized difference water in-dex (NDWI) is an important metric, which takes advantageof the fact that the reflectance of water areas is near to zeroin near-infrared (NIR) band and high in green band. Lots ofworks have been done on sea–land segmentation and coast-line extraction for multispectral images based on the NDWIanalysis .For panchromatic and natural-colored images, directly us-ing global threshold often fails due to the complicated dis-tribution of intensity and texture in land and sea areas.
Segmentation results of different methods. (a) Origi-nal image. (b) Result of Otsu. (c) Result of method in .(d) Result of our method. (e) Ground truth. Both (b) and (c)contain misclassification in land regions. Our method obtainsspatially consistent results by combining region merging and graph cut.
For most sea–land images, the sea areas present an even distri-bution of intensity and regular distribution of texture. Whilefor land areas, the situations are complicated due to variousobjects on the ground. Thus, we can utilize the area of homo-geneous regions as a metric to differentiate sea from land. Tothis end, a two-level region merging method is proposed.On the first level, neighboring homogeneous pixels aregrouped into superpixels by SLIC [13] method. The super-pixels are calculated in the[l,a,b,x,y]Tfeature space, wherel,aandbare the three channels in CIELAB color space. Weset the size of superpixels at 50 and use them to construct thegraph models for region merging and GC.We encode the superpixels by calculating the morpho-logical profiles (MPs) and morphological attribute profiles(APs) [14], which are effective methods for extracting spa-tial characteristics of objects in remote sensing images. TheMPs and APs of the original image are calculated in RGBchannels and the feature vector of pixeliis defined aszi=(sTi,mpTi,apTi)T, wheresTi,mpTiandapTiare thecolor, MPs and APs vector ofi. Each superpixel is encodedwith the mean value ofziof the pixels within it. The lengthof the feature vector is 84