Road extraction from high-resolution remote sensing images (HRSI) is a challenging but hot research topic in the past decades. A large number of methods are invented to deal with this problem. This article provides a comprehensive review of these existing approaches. We classified the methods into heuristic and data-driven. The heuristic methods are the mainstream in the early years, and the data-driven methods based on deep learning have been quickly developed recently. With regard to the heuristic methods, the road feature model is firstly introduced, then, the classic extraction methods are reviewed in two sub-categories: semi-automatic and automatic. The principles, inspirations, advantages and disadvantages of these methods are described. In terms of the data-driven methods, the road extraction methods based on deep neural network, particularly those based on patched convolutional neural network, fully convolutional network and generative adversarial network are reviewed. We perform subjective comparisons between the methods inner each type. Furthermore, the quantity performances achieved on the same dataset are compared between the heuristic and data-driven methods to show the strengthening of the data-driven methods. Finally, the conclusion and prospects are summarized.

The road center-line extraction is the key step of the road network extraction and modeling. The hand-craft feature engineering in the traditional road extraction methods is unstable which makes the extracted road center-line deviated from the road center in complex cases and even results in overall extracting errors. Recently, the road center-line extraction methods based on semantic segmentation employing deep neural network greatly outperformed the traditional methods. Nevertheless, the pixel-wise labels for training deep learning models are expensive and the post-process of road segmentation is error-prone. Inspired by the work of human pose estimation, we propose DeepWindow, a novel method to automatically extract the road network from remote sensing images. DeepWindow uses a sliding window guided by a CNN-based decision function to track the road network directly from the images without the prior of road segmentation. First of all, we design and train a CNN model to estimate the road center points inside a patch. Then, the road seeds are automatically searched patch by patch employing the CNN model. Finally, starting from seeds, our method firstly estimates the road direction using a Fourier spectrum analysis algorithm and then iteratively tracking the road center-line along the road direction guided by the CNN model. In our method, the CNN model is trained by point annotations, which greatly reduces the training costs comparing to those in semantic model training. Our method achieves comparable performance with the state-of-the-art road extraction methods, extensive experiments indicate that our method is robust to the point deviation.

Road segmentation methods based on deep neural networks have achieved great success in recent years, but creating accurate pixel-wise training labels is still a boring and expensive task, especially for large-scale high-resolution remote sensing images (HRSIs). Inspired by the stacked hourglass model for human joints detection, we propose a weakly supervised road segmentation method using point annotations in this article. First, we design a patch-based deep convolutional neural network (DCNN) model for road seeds and background points detection and train the model using point annotations. Then, in the process of road segmentation, the DCNN model detects a series of road and background points that are used to train a Support Vector Machine Classifier (SVC) for classifying each pixel into road or nonroad. According to the local geometry of road and the inaccurate classification of SVC, a multiscale and multidirection Gabor filter (MMGF) is put forward to estimate the road potential. Finally, the active contour model based on local binary fitting energy (LBF-Snake) is introduced to extract the road regions from the inhomogeneous road potential. Qualitative and quantitative comparisons show that our method achieves results close to the fully supervised semantic methods without considering the annotation cost and outperforms them given a fixed budget.

Road digitizing is a labor-intensive task. For reducing labor cost, this letter presents an effective semiautomatic road delineation method based on geodesic distance field (GDF) and piecewise polygon fitting. The main components of the approach are optimal circle calculation, soft road center kernel density estimation (KDE), fast GDF generation, and piecewise polygon fitting model. First, the adaptive circular template is proposed to automatically measure the road width. Next, the soft road center kernel density is estimated for fast GDF generation, which supports the extraction of road centerline between two adjacent seeds. Finally, piecewise polygon fitting is used to refine the road centerline. Extensive experiments demonstrate that the proposed algorithm is efficient and robust in road delineation. The proposed approach takes almost the same time to extract any length of road segment given fixed image size, and no hyperparameters needs to be set.