Unsupervised road extraction via a Gaussian mixture model with object-based features

ABSTRACT Automatic road extraction from remotely sensed images is an important and challenging task. This article proposes an unsupervised road detection method based on a Gaussian mixture model and object-based features. Our approach has five major stages, i.e. superpixel segmentation, feature description, homogeneous region merging, clustering via the Gaussian mixture model, and outlier filtering. In the third step, we present a graph-based region merging algorithm, in which the nodes of the graph are superpixels and edges are the similarities of intensity, colour, and texture. We also define two shape features, called deviation of parallelism (DoP) and narrow rate (NR), to automatically recognize road layer and filter outliers in the last step. We evaluated the proposed method on a variety of datasets, in which the Vaihingen dataset from the International Society for Photogrammetry and Remote Sensing Test Project is also included. Results demonstrate the power of our approach compared with some state-of-the-art methods.