Research On Semi-automatic Extraction Of Landslide From Remote Sensing Image Based On Digital Matting

Landslide is a phenomenon that the unstable rock and soil fall down along sliding surface under gravity. It usually leads to loss more industrial and agricultural production and human's life and property, even devastation. It is interesting to research that how to extract the landslide from remote sensing images efficiently and accurately. Landslides extraction is a technology to identify and extract the landslide from remote sensing images and separate the landslide target from background. The digital matting is a technology to separate a certain part of the picture or video into 2 layers and prepare for the post-image analysis. Use a few part of foreground and background specified by user to extract the foreground according to certain rules. This paper applies the digital matting to landslide extraction, presents landslide extraction method based on sample matting, landslide extraction method based on spectral matting, landslide extraction method based on mixed matting.1. Sample matting is often used to extract landslide area. The user usually separates the image into foreground region, background region and the unknown region. It is use foreground regions and background regions to estimate the foreground and background of unknown regions and get the alpha, then extract the landslide from the image.2. In the second method, first of all, hill-climbing algorithm is used to segment the image. Then spectral matting technology is used to get the foreground transparency. Lastly foreground-reconstructing algorithm is used to extract the target. Because of three problems when using spectral matting method directly to extract landslide from remote sensing image, we improve the traditional methods in the following three aspects:(1) Use self-adaptive method to choose the best number of clusters automatically to extract landslide.(2) Use the high-climbing algorithm for image segmentation, find the highest point among the local color histogram to cluster the adjacent pixels.(3) Remove the smooth items of matting component, and project the clustering results indicate vector onto the feature vectors directly as a matting component.3. In the third method, combining the first and the second method, then the Gibbs energy function is established. This function includes data item and smooth item, the former item estimates foreground and background by sample matting. The later item smoothes the edges of image, and minimizes the energy function. Finally the foreground and background can be separated to extract the landslide.Theoretical analysis and experimental results show that the samples matting, mixed matting, matting landslide spectrum extraction method in remote sensing image of the extracted landslide achieved good results, high accuracy and stability.