|
In recent years, remote sensing imaging technology has developed rapidly. A growing number of high resolution remote sensing images become available, which largely facilitates the research and applications of remote sensing images. Landcover classification is one of the most important tasks of remote sensing image applications [1]. However, traditional classification methods rely on manual feature design, which is time-consuming and requires expertise. It is difficult to apply to massive data. Compared with the traditional classification methods, deep learning [2] can automatically acquire the most intrinsic and discriminative features of the image. Based on the deep learning image classification, this paper designs a high-level semantic information extraction system with high efficiency and robustness. A deep fully convolutional networks (FCN) is designed to extract the features from remote sensing images and to predict the landcover classes of each image, which include building, tree, road, and grass. On the basis of the classification results, we use binarization to highlight the building objects. Then the noise of the binarized image is removed by Gaussian filtering and morphological image processing. After that we set a threshold to delete small misdiagnosis areas. At last the connected domain algorithm is applied to detect the buildings and calculate the building number in each image. The forest coverage is then obtained by computing the proportion of the pixels with ‘tree’ class label to the total number of the pixels in each image. Different from the traditional image interpretation method, this systematic high-level semantic information extraction framework not only detects the number of buildings in the scene but also extracts forest coverage. Moreover, more high-level information extraction can be easily supplemented to this framework, such as road localization or interested object detection.
|
