High-resolution Remote Sensing Image Water-land Separation Algorithm Based On Texture Feature

With the development of high resolution remote sensing image imaging technology, target recognition and information extraction based on high resolution remote sensing image has become a hot spot. Separation of water and land is of great significance in offshore monitoring and harbor ship target recognition. The image texture feature is one of the most important features of image. Image texture not only reflects the grayscale statistics of the image and the spatial distribution of grey information, but also reflects the local details of the image. So image texture information is a significant feature to distinct water and land in the high resolution remote sensing images.Based on the analysis of the port area of numerous high resolution remote sensing images, namely, the texture of water region is more smooth and the texture of land region changes largely, a high resolution remote sensing image water-land separation method is put forward based on texture feature.First of all, this paper introduces several kinds of image texture feature extracting methods, and analyzes their performance. Afterwards, the method with a good performance is selected to extract texture feature as the final approach.Then, the images of water background and land background are labeled artificially as the training samples to extract texture feature respectively and a support vector machine(SVM) classifier is trained for the texture feature. And the test images which are to be classified are split into sub-block segmentations. Then each block of the image is judged by the built classifier to decide which class it belongs to. The classification results are marked as image coarse segmentation results.Finally, the image coarse segmentation results are optimized by filtering of connected components after the binarization while the mistaken components are removed. The connected components of the binary images which transform from image coarse segmentation results are labeled and analyzed to selected a threshold. Then the component which is smaller than the threshold area is removed. Then remote sensing image water-land separation can be selected. In order to separate images with higher accuracy, the boundary areas of the images are split into smaller sub-block segmentations and a texture feature classifier is built for the boundary areas of the images to classify the segmentations again. At last, the classification results are considered as the final water-land separation results.On the basis of the water-land separation results, methods which can roughly detect targets in the port remote sensing images are proposed. The offshore targets can be detected by filtering connected area of the coarse segmentation results. As for the targets docked in the images, a newly built texture feature classifier is trained by the texture feature of the targets docked and the texture feature of the others. Then the area which is to be classified is split into smaller segmentations and the segmentations are classified one by one. Finally, the targets docked can be detected.