Road information derived from remote sensing images is extensively utilized across various fields. Traditionally, road extraction methods relied predominantly on manually designed features, and yet these have progressively evolved towards deep learning-based approaches. Despite robust performances by existing methodologies, extracting road regions from high-resolution remote sensing images still presents significant challenges: 1) the narrow and continuous morphological characteristics of roads require comprehensive contextual understanding from a global perspective, and 2) roads are often obscured by vegetation, shadows, and buildings, causing visual discontinuities and a lack of local correlation. To address these issues, this paper proposes a hierarchical feature matching network for separated features of structure-semantic. It is mainly composed of the Structure-Semantic Features Enhancement Pipeline (SFEP) with two parallel modules combined a fusion module (SFMF). This separated architecture allows each pattern of features to maintain distinct representations without mutual interference, improving inter-regional correlation. Meanwhile, the SFMF is used to match and fuse the diverse patterns of features for the corresponding regions. Additionally, a novel Connected Region (CR) loss is proposed to leverage the inherent topological connectivity of roads to direct the attention of model to truncated regions. Experiments are conducted on the DeepGlobe, CITY-OSM and CHN6-CUG datasets, and the results show that our method significantly outperforms existing networks on most evaluation metrics, highlighting its effectiveness in addressing complex road extraction challenges.
The use of remote sensing images and supplemented by deep learning methods for road extraction has greatly accelerated work efficiency while still maintaining remarkable accuracy. In recent years, researchers have attempted to adopt a collaborative design of multiple interrelated tasks to further improve the performance of road extraction. Although the design expectations are met in most cases, there are still some problems: 1) The interaction among tasks is often only related to shared parameters or feature fusion with poor interpretability, and 2) There may be conflicts among tasks, which could lead to a decline in performance instead of an increase. To solve the above problems, this article proposes a novel hybrid network. This network adopts the integration of different architectures, each efficiently demonstrating its ability to extract global and local features, and achieving complementarity among them. Furthermore, we proposed the DBLFR module for guiding feature refinement from the perspective of the frequency domain. In addition, compared with implicit multitask interaction, the proposal of EMTL attempts to regard the multitask structure as a deeply supervised process and improve the network’s attention to the main tasks through explicit fusion. Our experiments on the DeepGlobe and CITY-OSM datasets verified the effectiveness of the proposed network and modules, and were significantly superior to the existing methods in most metrics.
Extracting road information from remote sensing imagery is essential for numerous applications. Traditional methods have heavily depended on manual feature extraction, but recent progress has shifted the focus towards deep learning-based approaches. However, many existing methods primarily focus on the continuity of local roads, neglecting challenges associated with road shape extraction. These challenges include interference from shadows, buildings, vehicles, and complex backgrounds, often resulting in discontinuous results. To tackle these issues, this paper introduces a novel U-shaped deep learning network that integrates a Hybrid Convolution Module and Block Loop Attention to achieve robust road extraction, which called HLU-Net. The Hybrid Convolution Module, with its two parallel branches featuring convolution kernels of varying scales, not only broadens the receptive field to capture comprehensive contextual information but also concentrates on local details for accurate feature extraction. Furthermore, each branch employs unique convolution techniques to reduce computational redundancy and enhance operational efficiency. In addition, the proposed Block Loop Attention mechanism effectively models the relationships between elements within feature vectors, capturing intrinsic dependencies, and thus facilitating more efficient computations without merely reducing complexity. Experiments conducted on the DeepGlobe and CITY-OSM datasets demonstrate that our approach generally surpasses several state-of-the-art road extraction methods.